feat(llm-gateway): implement LLM Gateway MCP Server (#56)

Implements complete LLM Gateway MCP Server with: - FastMCP server with 4 tools: chat_completion, list_models, get_usage, count_tokens - LiteLLM Router with multi-provider failover chains - Circuit breaker pattern for fault tolerance - Redis-based cost tracking per project/agent - Comprehensive test suite (209 tests, 92% coverage) Model groups defined per ADR-004: - reasoning: claude-opus-4 → gpt-4.1 → gemini-2.5-pro - code: claude-sonnet-4 → gpt-4.1 → deepseek-coder - fast: claude-haiku → gpt-4.1-mini → gemini-2.0-flash 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-01-03 20:31:19 +01:00
parent 746fb7b181
commit 6e8b0b022a
23 changed files with 9794 additions and 93 deletions
--- a/mcp-servers/llm-gateway/Dockerfile
+++ b/mcp-servers/llm-gateway/Dockerfile
@@ -0,0 +1,53 @@
+# Syndarix LLM Gateway MCP Server
+# Multi-stage build for minimal image size
+
+# Build stage
+FROM python:3.12-slim AS builder
+
+# Install uv for fast package management
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+
+WORKDIR /app
+
+# Copy dependency files
+COPY pyproject.toml ./
+
+# Create virtual environment and install dependencies
+RUN uv venv /app/.venv
+ENV PATH="/app/.venv/bin:$PATH"
+RUN uv pip install -e .
+
+# Runtime stage
+FROM python:3.12-slim AS runtime
+
+# Create non-root user for security
+RUN groupadd --gid 1000 appgroup && \
+    useradd --uid 1000 --gid appgroup --shell /bin/bash --create-home appuser
+
+WORKDIR /app
+
+# Copy virtual environment from builder
+COPY --from=builder /app/.venv /app/.venv
+ENV PATH="/app/.venv/bin:$PATH"
+
+# Copy application code
+COPY --chown=appuser:appgroup . .
+
+# Switch to non-root user
+USER appuser
+
+# Environment variables
+ENV LLM_GATEWAY_HOST=0.0.0.0
+ENV LLM_GATEWAY_PORT=8001
+ENV PYTHONUNBUFFERED=1
+ENV PYTHONDONTWRITEBYTECODE=1
+
+# Expose port
+EXPOSE 8001
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
+    CMD python -c "import httpx; httpx.get('http://localhost:8001/health').raise_for_status()" || exit 1
+
+# Run the server
+CMD ["python", "server.py"]
--- a/mcp-servers/llm-gateway/config.py
+++ b/mcp-servers/llm-gateway/config.py
@@ -0,0 +1,179 @@
+"""
+Configuration for LLM Gateway MCP Server.
+
+Uses Pydantic Settings for type-safe environment variable handling.
+"""
+
+from functools import lru_cache
+
+from pydantic import Field, field_validator
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+    """LLM Gateway configuration settings."""
+
+    model_config = SettingsConfigDict(
+        env_prefix="LLM_GATEWAY_",
+        env_file=".env",
+        env_file_encoding="utf-8",
+        extra="ignore",
+    )
+
+    # Server settings
+    host: str = Field(default="0.0.0.0", description="Server host")
+    port: int = Field(default=8001, description="Server port")
+    debug: bool = Field(default=False, description="Debug mode")
+
+    # Redis settings
+    redis_url: str = Field(
+        default="redis://localhost:6379/0",
+        description="Redis connection URL",
+    )
+    redis_prefix: str = Field(
+        default="llm_gateway",
+        description="Redis key prefix",
+    )
+    redis_ttl_hours: int = Field(
+        default=24,
+        description="Default Redis TTL in hours",
+    )
+
+    # Provider API keys
+    anthropic_api_key: str | None = Field(
+        default=None,
+        description="Anthropic API key for Claude models",
+    )
+    openai_api_key: str | None = Field(
+        default=None,
+        description="OpenAI API key for GPT models",
+    )
+    google_api_key: str | None = Field(
+        default=None,
+        description="Google API key for Gemini models",
+    )
+    alibaba_api_key: str | None = Field(
+        default=None,
+        description="Alibaba API key for Qwen models",
+    )
+    deepseek_api_key: str | None = Field(
+        default=None,
+        description="DeepSeek API key",
+    )
+    deepseek_base_url: str | None = Field(
+        default=None,
+        description="DeepSeek API base URL (for self-hosted)",
+    )
+
+    # LiteLLM settings
+    litellm_timeout: int = Field(
+        default=120,
+        description="LiteLLM request timeout in seconds",
+    )
+    litellm_max_retries: int = Field(
+        default=3,
+        description="Maximum retries per provider",
+    )
+    litellm_cache_enabled: bool = Field(
+        default=True,
+        description="Enable Redis caching for LiteLLM",
+    )
+    litellm_cache_ttl: int = Field(
+        default=3600,
+        description="Cache TTL in seconds",
+    )
+
+    # Circuit breaker settings
+    circuit_failure_threshold: int = Field(
+        default=5,
+        description="Failures before circuit opens",
+    )
+    circuit_recovery_timeout: int = Field(
+        default=60,
+        description="Seconds before circuit half-opens",
+    )
+    circuit_half_open_max_calls: int = Field(
+        default=3,
+        description="Max calls in half-open state",
+    )
+
+    # Cost tracking settings
+    cost_tracking_enabled: bool = Field(
+        default=True,
+        description="Enable cost tracking",
+    )
+    cost_alert_threshold: float = Field(
+        default=100.0,
+        description="Cost threshold for alerts (USD)",
+    )
+    default_budget_limit: float = Field(
+        default=1000.0,
+        description="Default project budget limit (USD)",
+    )
+
+    # Rate limiting
+    rate_limit_enabled: bool = Field(
+        default=True,
+        description="Enable rate limiting",
+    )
+    rate_limit_requests_per_minute: int = Field(
+        default=60,
+        description="Max requests per minute per project",
+    )
+
+    @field_validator("port")
+    @classmethod
+    def validate_port(cls, v: int) -> int:
+        """Validate port is in valid range."""
+        if not 1 <= v <= 65535:
+            raise ValueError("Port must be between 1 and 65535")
+        return v
+
+    @field_validator("redis_ttl_hours")
+    @classmethod
+    def validate_ttl(cls, v: int) -> int:
+        """Validate TTL is positive."""
+        if v <= 0:
+            raise ValueError("Redis TTL must be positive")
+        return v
+
+    @field_validator("circuit_failure_threshold")
+    @classmethod
+    def validate_failure_threshold(cls, v: int) -> int:
+        """Validate failure threshold is reasonable."""
+        if not 1 <= v <= 100:
+            raise ValueError("Failure threshold must be between 1 and 100")
+        return v
+
+    @field_validator("litellm_timeout")
+    @classmethod
+    def validate_timeout(cls, v: int) -> int:
+        """Validate timeout is reasonable."""
+        if not 1 <= v <= 600:
+            raise ValueError("Timeout must be between 1 and 600 seconds")
+        return v
+
+    def get_available_providers(self) -> list[str]:
+        """Get list of providers with configured API keys."""
+        providers = []
+        if self.anthropic_api_key:
+            providers.append("anthropic")
+        if self.openai_api_key:
+            providers.append("openai")
+        if self.google_api_key:
+            providers.append("google")
+        if self.alibaba_api_key:
+            providers.append("alibaba")
+        if self.deepseek_api_key or self.deepseek_base_url:
+            providers.append("deepseek")
+        return providers
+
+    def has_any_provider(self) -> bool:
+        """Check if at least one provider is configured."""
+        return len(self.get_available_providers()) > 0
+
+
+@lru_cache
+def get_settings() -> Settings:
+    """Get cached settings instance."""
+    return Settings()
--- a/mcp-servers/llm-gateway/cost_tracking.py
+++ b/mcp-servers/llm-gateway/cost_tracking.py
@@ -0,0 +1,467 @@
+"""
+Cost tracking for LLM Gateway.
+
+Tracks LLM usage costs per project and agent using Redis.
+Provides aggregation by hour, day, and month with TTL-based expiry.
+"""
+
+import logging
+from datetime import UTC, datetime, timedelta
+from typing import Any
+
+import redis.asyncio as redis
+
+from config import Settings, get_settings
+from models import (
+    MODEL_CONFIGS,
+    UsageReport,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class CostTracker:
+    """
+    Redis-based cost tracker for LLM usage.
+
+    Key structure:
+    - {prefix}:cost:project:{project_id}:{date} -> Hash of usage by model
+    - {prefix}:cost:agent:{agent_id}:{date} -> Hash of usage by model
+    - {prefix}:cost:session:{session_id} -> Hash of session usage
+    - {prefix}:requests:{project_id}:{date} -> Request count
+
+    Date formats:
+    - hour: YYYYMMDDHH
+    - day: YYYYMMDD
+    - month: YYYYMM
+    """
+
+    def __init__(
+        self,
+        redis_client: redis.Redis | None = None,
+        settings: Settings | None = None,
+    ) -> None:
+        """
+        Initialize cost tracker.
+
+        Args:
+            redis_client: Redis client (creates one if None)
+            settings: Application settings
+        """
+        self._settings = settings or get_settings()
+        self._redis: redis.Redis | None = redis_client
+        self._prefix = self._settings.redis_prefix
+
+    async def _get_redis(self) -> redis.Redis:
+        """Get or create Redis client."""
+        if self._redis is None:
+            self._redis = redis.from_url(
+                self._settings.redis_url,
+                decode_responses=True,
+            )
+        return self._redis
+
+    async def close(self) -> None:
+        """Close Redis connection."""
+        if self._redis:
+            await self._redis.aclose()
+            self._redis = None
+
+    def _get_date_keys(self, timestamp: datetime | None = None) -> dict[str, str]:
+        """Get date format keys for different periods."""
+        if timestamp is None:
+            timestamp = datetime.now(UTC)
+        return {
+            "hour": timestamp.strftime("%Y%m%d%H"),
+            "day": timestamp.strftime("%Y%m%d"),
+            "month": timestamp.strftime("%Y%m"),
+        }
+
+    def _get_ttl_seconds(self, period: str) -> int:
+        """Get TTL in seconds for a period."""
+        ttls = {
+            "hour": 24 * 3600,  # 24 hours
+            "day": 30 * 24 * 3600,  # 30 days
+            "month": 365 * 24 * 3600,  # 1 year
+        }
+        return ttls.get(period, 30 * 24 * 3600)
+
+    async def record_usage(
+        self,
+        project_id: str,
+        agent_id: str,
+        model: str,
+        prompt_tokens: int,
+        completion_tokens: int,
+        cost_usd: float,
+        session_id: str | None = None,
+        request_id: str | None = None,  # noqa: ARG002 - reserved for future logging
+    ) -> None:
+        """
+        Record LLM usage.
+
+        Args:
+            project_id: Project ID
+            agent_id: Agent ID
+            model: Model name
+            prompt_tokens: Input tokens
+            completion_tokens: Output tokens
+            cost_usd: Cost in USD
+            session_id: Optional session ID
+            request_id: Optional request ID
+        """
+        if not self._settings.cost_tracking_enabled:
+            return
+
+        r = await self._get_redis()
+        date_keys = self._get_date_keys()
+        pipe = r.pipeline()
+
+        # Record for each time period
+        for period, date_key in date_keys.items():
+            # Project-level tracking
+            project_key = f"{self._prefix}:cost:project:{project_id}:{date_key}"
+            await self._increment_usage(
+                pipe, project_key, model, prompt_tokens, completion_tokens, cost_usd
+            )
+            pipe.expire(project_key, self._get_ttl_seconds(period))
+
+            # Agent-level tracking
+            agent_key = f"{self._prefix}:cost:agent:{agent_id}:{date_key}"
+            await self._increment_usage(
+                pipe, agent_key, model, prompt_tokens, completion_tokens, cost_usd
+            )
+            pipe.expire(agent_key, self._get_ttl_seconds(period))
+
+            # Request counter
+            requests_key = f"{self._prefix}:requests:{project_id}:{date_key}"
+            pipe.incr(requests_key)
+            pipe.expire(requests_key, self._get_ttl_seconds(period))
+
+        # Session tracking (if session_id provided)
+        if session_id:
+            session_key = f"{self._prefix}:cost:session:{session_id}"
+            await self._increment_usage(
+                pipe, session_key, model, prompt_tokens, completion_tokens, cost_usd
+            )
+            pipe.expire(session_key, 24 * 3600)  # 24 hour TTL for sessions
+
+        await pipe.execute()
+
+        logger.debug(
+            f"Recorded usage: project={project_id}, agent={agent_id}, "
+            f"model={model}, tokens={prompt_tokens + completion_tokens}, "
+            f"cost=${cost_usd:.6f}"
+        )
+
+    async def _increment_usage(
+        self,
+        pipe: redis.client.Pipeline,
+        key: str,
+        model: str,
+        prompt_tokens: int,
+        completion_tokens: int,
+        cost_usd: float,
+    ) -> None:
+        """Increment usage in a hash."""
+        # Store as JSON fields within the hash
+        pipe.hincrby(key, f"{model}:prompt_tokens", prompt_tokens)
+        pipe.hincrby(key, f"{model}:completion_tokens", completion_tokens)
+        pipe.hincrbyfloat(key, f"{model}:cost_usd", cost_usd)
+        pipe.hincrby(key, f"{model}:requests", 1)
+        # Totals
+        pipe.hincrby(key, "total:prompt_tokens", prompt_tokens)
+        pipe.hincrby(key, "total:completion_tokens", completion_tokens)
+        pipe.hincrbyfloat(key, "total:cost_usd", cost_usd)
+        pipe.hincrby(key, "total:requests", 1)
+
+    async def get_project_usage(
+        self,
+        project_id: str,
+        period: str = "day",
+        timestamp: datetime | None = None,
+    ) -> UsageReport:
+        """
+        Get usage report for a project.
+
+        Args:
+            project_id: Project ID
+            period: Time period (hour, day, month)
+            timestamp: Specific time to query (defaults to now)
+
+        Returns:
+            Usage report
+        """
+        date_keys = self._get_date_keys(timestamp)
+        date_key = date_keys.get(period, date_keys["day"])
+
+        key = f"{self._prefix}:cost:project:{project_id}:{date_key}"
+        return await self._get_usage_report(
+            key, project_id, "project", period, timestamp
+        )
+
+    async def get_agent_usage(
+        self,
+        agent_id: str,
+        period: str = "day",
+        timestamp: datetime | None = None,
+    ) -> UsageReport:
+        """
+        Get usage report for an agent.
+
+        Args:
+            agent_id: Agent ID
+            period: Time period (hour, day, month)
+            timestamp: Specific time to query (defaults to now)
+
+        Returns:
+            Usage report
+        """
+        date_keys = self._get_date_keys(timestamp)
+        date_key = date_keys.get(period, date_keys["day"])
+
+        key = f"{self._prefix}:cost:agent:{agent_id}:{date_key}"
+        return await self._get_usage_report(key, agent_id, "agent", period, timestamp)
+
+    async def _get_usage_report(
+        self,
+        key: str,
+        entity_id: str,
+        entity_type: str,
+        period: str,
+        timestamp: datetime | None,
+    ) -> UsageReport:
+        """Get usage report from a Redis hash."""
+        r = await self._get_redis()
+        data = await r.hgetall(key)
+
+        # Parse the hash data
+        by_model: dict[str, dict[str, Any]] = {}
+        total_requests = 0
+        total_tokens = 0
+        total_cost = 0.0
+
+        for field, value in data.items():
+            parts = field.split(":")
+            if len(parts) != 2:
+                continue
+
+            model, metric = parts
+            if model == "total":
+                if metric == "requests":
+                    total_requests = int(value)
+                elif metric == "prompt_tokens" or metric == "completion_tokens":
+                    total_tokens += int(value)
+                elif metric == "cost_usd":
+                    total_cost = float(value)
+            else:
+                if model not in by_model:
+                    by_model[model] = {
+                        "prompt_tokens": 0,
+                        "completion_tokens": 0,
+                        "cost_usd": 0.0,
+                        "requests": 0,
+                    }
+                if metric == "prompt_tokens":
+                    by_model[model]["prompt_tokens"] = int(value)
+                elif metric == "completion_tokens":
+                    by_model[model]["completion_tokens"] = int(value)
+                elif metric == "cost_usd":
+                    by_model[model]["cost_usd"] = float(value)
+                elif metric == "requests":
+                    by_model[model]["requests"] = int(value)
+
+        # Calculate period boundaries
+        now = timestamp or datetime.now(UTC)
+        if period == "hour":
+            period_start = now.replace(minute=0, second=0, microsecond=0)
+            period_end = period_start + timedelta(hours=1)
+        elif period == "day":
+            period_start = now.replace(hour=0, minute=0, second=0, microsecond=0)
+            period_end = period_start + timedelta(days=1)
+        else:  # month
+            period_start = now.replace(day=1, hour=0, minute=0, second=0, microsecond=0)
+            # Next month
+            if now.month == 12:
+                period_end = period_start.replace(year=now.year + 1, month=1)
+            else:
+                period_end = period_start.replace(month=now.month + 1)
+
+        return UsageReport(
+            entity_id=entity_id,
+            entity_type=entity_type,
+            period=period,
+            period_start=period_start,
+            period_end=period_end,
+            total_requests=total_requests,
+            total_tokens=total_tokens,
+            total_cost_usd=round(total_cost, 6),
+            by_model=by_model,
+        )
+
+    async def get_session_usage(self, session_id: str) -> dict[str, Any]:
+        """
+        Get usage for a specific session.
+
+        Args:
+            session_id: Session ID
+
+        Returns:
+            Session usage data
+        """
+        r = await self._get_redis()
+        key = f"{self._prefix}:cost:session:{session_id}"
+        data = await r.hgetall(key)
+
+        # Parse similar to _get_usage_report
+        result: dict[str, Any] = {
+            "session_id": session_id,
+            "total_tokens": 0,
+            "total_cost_usd": 0.0,
+            "by_model": {},
+        }
+
+        for field, value in data.items():
+            parts = field.split(":")
+            if len(parts) != 2:
+                continue
+
+            model, metric = parts
+            if model == "total":
+                if metric == "prompt_tokens" or metric == "completion_tokens":
+                    result["total_tokens"] += int(value)
+                elif metric == "cost_usd":
+                    result["total_cost_usd"] = float(value)
+            else:
+                if model not in result["by_model"]:
+                    result["by_model"][model] = {}
+                if metric in ("prompt_tokens", "completion_tokens", "requests"):
+                    result["by_model"][model][metric] = int(value)
+                elif metric == "cost_usd":
+                    result["by_model"][model][metric] = float(value)
+
+        return result
+
+    async def check_budget(
+        self,
+        project_id: str,
+        budget_limit: float | None = None,
+    ) -> tuple[bool, float, float]:
+        """
+        Check if project is within budget.
+
+        Args:
+            project_id: Project ID
+            budget_limit: Budget limit (uses default if None)
+
+        Returns:
+            Tuple of (within_budget, current_cost, limit)
+        """
+        limit = budget_limit or self._settings.default_budget_limit
+
+        # Get current month usage
+        report = await self.get_project_usage(project_id, period="month")
+        current_cost = report.total_cost_usd
+
+        within_budget = current_cost < limit
+        return within_budget, current_cost, limit
+
+    async def estimate_request_cost(
+        self,
+        model: str,
+        prompt_tokens: int,
+        max_completion_tokens: int,
+    ) -> float:
+        """
+        Estimate cost for a request.
+
+        Args:
+            model: Model name
+            prompt_tokens: Input token count
+            max_completion_tokens: Maximum output tokens
+
+        Returns:
+            Estimated cost in USD
+        """
+        config = MODEL_CONFIGS.get(model)
+        if not config:
+            # Use a default estimate
+            return (prompt_tokens + max_completion_tokens) * 0.00001
+
+        input_cost = (prompt_tokens / 1_000_000) * config.cost_per_1m_input
+        output_cost = (max_completion_tokens / 1_000_000) * config.cost_per_1m_output
+        return round(input_cost + output_cost, 6)
+
+    async def should_alert(
+        self,
+        project_id: str,
+        threshold: float | None = None,
+    ) -> tuple[bool, float]:
+        """
+        Check if cost alert should be triggered.
+
+        Args:
+            project_id: Project ID
+            threshold: Alert threshold (uses default if None)
+
+        Returns:
+            Tuple of (should_alert, current_cost)
+        """
+        thresh = threshold or self._settings.cost_alert_threshold
+
+        report = await self.get_project_usage(project_id, period="day")
+        current_cost = report.total_cost_usd
+
+        return current_cost >= thresh, current_cost
+
+
+def calculate_cost(
+    model: str,
+    prompt_tokens: int,
+    completion_tokens: int,
+) -> float:
+    """
+    Calculate cost for a completion.
+
+    Args:
+        model: Model name
+        prompt_tokens: Input tokens
+        completion_tokens: Output tokens
+
+    Returns:
+        Cost in USD
+    """
+    config = MODEL_CONFIGS.get(model)
+    if not config:
+        logger.warning(f"Unknown model {model} for cost calculation")
+        return 0.0
+
+    input_cost = (prompt_tokens / 1_000_000) * config.cost_per_1m_input
+    output_cost = (completion_tokens / 1_000_000) * config.cost_per_1m_output
+    return round(input_cost + output_cost, 6)
+
+
+# Global tracker instance (lazy initialization)
+_tracker: CostTracker | None = None
+
+
+def get_cost_tracker() -> CostTracker:
+    """Get the global cost tracker instance."""
+    global _tracker
+    if _tracker is None:
+        _tracker = CostTracker()
+    return _tracker
+
+
+async def close_cost_tracker() -> None:
+    """Close the global cost tracker."""
+    global _tracker
+    if _tracker:
+        await _tracker.close()
+        _tracker = None
+
+
+def reset_cost_tracker() -> None:
+    """Reset the global tracker (for testing)."""
+    global _tracker
+    _tracker = None
--- a/mcp-servers/llm-gateway/exceptions.py
+++ b/mcp-servers/llm-gateway/exceptions.py
@@ -0,0 +1,478 @@
+"""
+Custom exceptions for LLM Gateway MCP Server.
+
+Provides structured error handling with error codes for consistent responses.
+"""
+
+from enum import Enum
+from typing import Any
+
+
+class ErrorCode(str, Enum):
+    """Error codes for LLM Gateway errors."""
+
+    # General errors
+    UNKNOWN_ERROR = "LLM_UNKNOWN_ERROR"
+    INVALID_REQUEST = "LLM_INVALID_REQUEST"
+    CONFIGURATION_ERROR = "LLM_CONFIGURATION_ERROR"
+
+    # Provider errors
+    PROVIDER_ERROR = "LLM_PROVIDER_ERROR"
+    PROVIDER_TIMEOUT = "LLM_PROVIDER_TIMEOUT"
+    PROVIDER_RATE_LIMIT = "LLM_PROVIDER_RATE_LIMIT"
+    PROVIDER_UNAVAILABLE = "LLM_PROVIDER_UNAVAILABLE"
+    ALL_PROVIDERS_FAILED = "LLM_ALL_PROVIDERS_FAILED"
+
+    # Model errors
+    INVALID_MODEL = "LLM_INVALID_MODEL"
+    INVALID_MODEL_GROUP = "LLM_INVALID_MODEL_GROUP"
+    MODEL_NOT_AVAILABLE = "LLM_MODEL_NOT_AVAILABLE"
+
+    # Circuit breaker errors
+    CIRCUIT_OPEN = "LLM_CIRCUIT_OPEN"
+    CIRCUIT_HALF_OPEN_EXHAUSTED = "LLM_CIRCUIT_HALF_OPEN_EXHAUSTED"
+
+    # Cost errors
+    COST_LIMIT_EXCEEDED = "LLM_COST_LIMIT_EXCEEDED"
+    BUDGET_EXHAUSTED = "LLM_BUDGET_EXHAUSTED"
+
+    # Rate limiting errors
+    RATE_LIMIT_EXCEEDED = "LLM_RATE_LIMIT_EXCEEDED"
+
+    # Streaming errors
+    STREAM_ERROR = "LLM_STREAM_ERROR"
+    STREAM_INTERRUPTED = "LLM_STREAM_INTERRUPTED"
+
+    # Token errors
+    TOKEN_LIMIT_EXCEEDED = "LLM_TOKEN_LIMIT_EXCEEDED"
+    CONTEXT_TOO_LONG = "LLM_CONTEXT_TOO_LONG"
+
+
+class LLMGatewayError(Exception):
+    """Base exception for LLM Gateway errors."""
+
+    def __init__(
+        self,
+        message: str,
+        code: ErrorCode = ErrorCode.UNKNOWN_ERROR,
+        details: dict[str, Any] | None = None,
+        cause: Exception | None = None,
+    ) -> None:
+        """
+        Initialize LLM Gateway error.
+
+        Args:
+            message: Human-readable error message
+            code: Error code for programmatic handling
+            details: Additional error details
+            cause: Original exception that caused this error
+        """
+        super().__init__(message)
+        self.message = message
+        self.code = code
+        self.details = details or {}
+        self.cause = cause
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert error to dictionary for JSON response."""
+        result = {
+            "error": self.code.value,
+            "message": self.message,
+        }
+        if self.details:
+            result["details"] = self.details
+        return result
+
+    def __str__(self) -> str:
+        """String representation."""
+        return f"[{self.code.value}] {self.message}"
+
+    def __repr__(self) -> str:
+        """Detailed representation."""
+        return (
+            f"{self.__class__.__name__}("
+            f"message={self.message!r}, "
+            f"code={self.code.value!r}, "
+            f"details={self.details!r})"
+        )
+
+
+class ProviderError(LLMGatewayError):
+    """Error from an LLM provider."""
+
+    def __init__(
+        self,
+        message: str,
+        provider: str,
+        model: str | None = None,
+        status_code: int | None = None,
+        details: dict[str, Any] | None = None,
+        cause: Exception | None = None,
+    ) -> None:
+        """
+        Initialize provider error.
+
+        Args:
+            message: Error message
+            provider: Provider that failed
+            model: Model that was being used
+            status_code: HTTP status code if applicable
+            details: Additional details
+            cause: Original exception
+        """
+        error_details = details or {}
+        error_details["provider"] = provider
+        if model:
+            error_details["model"] = model
+        if status_code:
+            error_details["status_code"] = status_code
+
+        super().__init__(
+            message=message,
+            code=ErrorCode.PROVIDER_ERROR,
+            details=error_details,
+            cause=cause,
+        )
+        self.provider = provider
+        self.model = model
+        self.status_code = status_code
+
+
+class RateLimitError(LLMGatewayError):
+    """Rate limit exceeded error."""
+
+    def __init__(
+        self,
+        message: str,
+        provider: str | None = None,
+        retry_after: int | None = None,
+        details: dict[str, Any] | None = None,
+    ) -> None:
+        """
+        Initialize rate limit error.
+
+        Args:
+            message: Error message
+            provider: Provider that rate limited (None for internal limit)
+            retry_after: Seconds until retry is allowed
+            details: Additional details
+        """
+        error_details = details or {}
+        if provider:
+            error_details["provider"] = provider
+        if retry_after:
+            error_details["retry_after_seconds"] = retry_after
+
+        code = (
+            ErrorCode.PROVIDER_RATE_LIMIT
+            if provider
+            else ErrorCode.RATE_LIMIT_EXCEEDED
+        )
+
+        super().__init__(
+            message=message,
+            code=code,
+            details=error_details,
+        )
+        self.provider = provider
+        self.retry_after = retry_after
+
+
+class CircuitOpenError(LLMGatewayError):
+    """Circuit breaker is open, provider temporarily unavailable."""
+
+    def __init__(
+        self,
+        provider: str,
+        recovery_time: int | None = None,
+        details: dict[str, Any] | None = None,
+    ) -> None:
+        """
+        Initialize circuit open error.
+
+        Args:
+            provider: Provider with open circuit
+            recovery_time: Seconds until circuit may recover
+            details: Additional details
+        """
+        error_details = details or {}
+        error_details["provider"] = provider
+        if recovery_time:
+            error_details["recovery_time_seconds"] = recovery_time
+
+        super().__init__(
+            message=f"Circuit breaker open for provider {provider}",
+            code=ErrorCode.CIRCUIT_OPEN,
+            details=error_details,
+        )
+        self.provider = provider
+        self.recovery_time = recovery_time
+
+
+class CostLimitExceededError(LLMGatewayError):
+    """Cost limit exceeded for project or agent."""
+
+    def __init__(
+        self,
+        entity_type: str,
+        entity_id: str,
+        current_cost: float,
+        limit: float,
+        details: dict[str, Any] | None = None,
+    ) -> None:
+        """
+        Initialize cost limit error.
+
+        Args:
+            entity_type: 'project' or 'agent'
+            entity_id: ID of the entity
+            current_cost: Current accumulated cost
+            limit: Cost limit that was exceeded
+            details: Additional details
+        """
+        error_details = details or {}
+        error_details["entity_type"] = entity_type
+        error_details["entity_id"] = entity_id
+        error_details["current_cost_usd"] = current_cost
+        error_details["limit_usd"] = limit
+
+        super().__init__(
+            message=(
+                f"Cost limit exceeded for {entity_type} {entity_id}: "
+                f"${current_cost:.2f} >= ${limit:.2f}"
+            ),
+            code=ErrorCode.COST_LIMIT_EXCEEDED,
+            details=error_details,
+        )
+        self.entity_type = entity_type
+        self.entity_id = entity_id
+        self.current_cost = current_cost
+        self.limit = limit
+
+
+class InvalidModelGroupError(LLMGatewayError):
+    """Invalid or unknown model group."""
+
+    def __init__(
+        self,
+        model_group: str,
+        available_groups: list[str] | None = None,
+    ) -> None:
+        """
+        Initialize invalid model group error.
+
+        Args:
+            model_group: The invalid group name
+            available_groups: List of valid group names
+        """
+        details: dict[str, Any] = {"requested_group": model_group}
+        if available_groups:
+            details["available_groups"] = available_groups
+
+        super().__init__(
+            message=f"Invalid model group: {model_group}",
+            code=ErrorCode.INVALID_MODEL_GROUP,
+            details=details,
+        )
+        self.model_group = model_group
+        self.available_groups = available_groups
+
+
+class InvalidModelError(LLMGatewayError):
+    """Invalid or unknown model."""
+
+    def __init__(
+        self,
+        model: str,
+        reason: str | None = None,
+    ) -> None:
+        """
+        Initialize invalid model error.
+
+        Args:
+            model: The invalid model name
+            reason: Reason why it's invalid
+        """
+        details: dict[str, Any] = {"requested_model": model}
+        if reason:
+            details["reason"] = reason
+
+        super().__init__(
+            message=f"Invalid model: {model}" + (f" ({reason})" if reason else ""),
+            code=ErrorCode.INVALID_MODEL,
+            details=details,
+        )
+        self.model = model
+
+
+class ModelNotAvailableError(LLMGatewayError):
+    """Model not available (provider not configured)."""
+
+    def __init__(
+        self,
+        model: str,
+        provider: str,
+    ) -> None:
+        """
+        Initialize model not available error.
+
+        Args:
+            model: The unavailable model
+            provider: The provider that's not configured
+        """
+        super().__init__(
+            message=f"Model {model} not available: {provider} provider not configured",
+            code=ErrorCode.MODEL_NOT_AVAILABLE,
+            details={"model": model, "provider": provider},
+        )
+        self.model = model
+        self.provider = provider
+
+
+class AllProvidersFailedError(LLMGatewayError):
+    """All providers in the failover chain failed."""
+
+    def __init__(
+        self,
+        model_group: str,
+        attempted_models: list[str],
+        errors: list[dict[str, Any]],
+    ) -> None:
+        """
+        Initialize all providers failed error.
+
+        Args:
+            model_group: The model group that was requested
+            attempted_models: Models that were attempted
+            errors: Errors from each attempt
+        """
+        super().__init__(
+            message=f"All providers failed for model group {model_group}",
+            code=ErrorCode.ALL_PROVIDERS_FAILED,
+            details={
+                "model_group": model_group,
+                "attempted_models": attempted_models,
+                "errors": errors,
+            },
+        )
+        self.model_group = model_group
+        self.attempted_models = attempted_models
+        self.errors = errors
+
+
+class StreamError(LLMGatewayError):
+    """Error during streaming response."""
+
+    def __init__(
+        self,
+        message: str,
+        chunks_received: int = 0,
+        cause: Exception | None = None,
+    ) -> None:
+        """
+        Initialize stream error.
+
+        Args:
+            message: Error message
+            chunks_received: Number of chunks received before error
+            cause: Original exception
+        """
+        super().__init__(
+            message=message,
+            code=ErrorCode.STREAM_ERROR,
+            details={"chunks_received": chunks_received},
+            cause=cause,
+        )
+        self.chunks_received = chunks_received
+
+
+class TokenLimitExceededError(LLMGatewayError):
+    """Request exceeds model's token limit."""
+
+    def __init__(
+        self,
+        model: str,
+        token_count: int,
+        limit: int,
+    ) -> None:
+        """
+        Initialize token limit error.
+
+        Args:
+            model: Model name
+            token_count: Requested token count
+            limit: Model's token limit
+        """
+        super().__init__(
+            message=f"Token count {token_count} exceeds {model} limit of {limit}",
+            code=ErrorCode.TOKEN_LIMIT_EXCEEDED,
+            details={
+                "model": model,
+                "requested_tokens": token_count,
+                "limit": limit,
+            },
+        )
+        self.model = model
+        self.token_count = token_count
+        self.limit = limit
+
+
+class ContextTooLongError(LLMGatewayError):
+    """Input context exceeds model's context window."""
+
+    def __init__(
+        self,
+        model: str,
+        context_length: int,
+        max_context: int,
+    ) -> None:
+        """
+        Initialize context too long error.
+
+        Args:
+            model: Model name
+            context_length: Input context length
+            max_context: Model's max context window
+        """
+        super().__init__(
+            message=(
+                f"Context length {context_length} exceeds {model} "
+                f"context window of {max_context}"
+            ),
+            code=ErrorCode.CONTEXT_TOO_LONG,
+            details={
+                "model": model,
+                "context_length": context_length,
+                "max_context": max_context,
+            },
+        )
+        self.model = model
+        self.context_length = context_length
+        self.max_context = max_context
+
+
+class ConfigurationError(LLMGatewayError):
+    """Configuration error."""
+
+    def __init__(
+        self,
+        message: str,
+        config_key: str | None = None,
+    ) -> None:
+        """
+        Initialize configuration error.
+
+        Args:
+            message: Error message
+            config_key: Configuration key that's problematic
+        """
+        details: dict[str, Any] = {}
+        if config_key:
+            details["config_key"] = config_key
+
+        super().__init__(
+            message=message,
+            code=ErrorCode.CONFIGURATION_ERROR,
+            details=details,
+        )
+        self.config_key = config_key
--- a/mcp-servers/llm-gateway/failover.py
+++ b/mcp-servers/llm-gateway/failover.py
@@ -0,0 +1,357 @@
+"""
+Circuit Breaker implementation for LLM Gateway.
+
+Provides fault tolerance by tracking provider failures and
+temporarily disabling providers that are experiencing issues.
+"""
+
+import asyncio
+import logging
+import time
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import Any, TypeVar
+
+from config import Settings, get_settings
+from exceptions import CircuitOpenError
+
+logger = logging.getLogger(__name__)
+
+T = TypeVar("T")
+
+
+class CircuitState(str, Enum):
+    """Circuit breaker states."""
+
+    CLOSED = "closed"  # Normal operation, requests pass through
+    OPEN = "open"  # Failures exceeded threshold, requests blocked
+    HALF_OPEN = "half_open"  # Testing if service recovered
+
+
+@dataclass
+class CircuitStats:
+    """Statistics for a circuit breaker."""
+
+    failures: int = 0
+    successes: int = 0
+    last_failure_time: float | None = None
+    last_success_time: float | None = None
+    state_changed_at: float = field(default_factory=time.time)
+    half_open_calls: int = 0
+
+
+class CircuitBreaker:
+    """
+    Circuit breaker for individual providers.
+
+    States:
+    - CLOSED: Normal operation. Failures increment counter.
+    - OPEN: Too many failures. Requests immediately fail.
+    - HALF_OPEN: Testing recovery. Limited requests allowed.
+
+    Transitions:
+    - CLOSED -> OPEN: When failures >= threshold
+    - OPEN -> HALF_OPEN: After recovery_timeout
+    - HALF_OPEN -> CLOSED: On success
+    - HALF_OPEN -> OPEN: On failure
+    """
+
+    def __init__(
+        self,
+        name: str,
+        failure_threshold: int = 5,
+        recovery_timeout: int = 60,
+        half_open_max_calls: int = 3,
+    ) -> None:
+        """
+        Initialize circuit breaker.
+
+        Args:
+            name: Identifier for this circuit (usually provider name)
+            failure_threshold: Failures before opening circuit
+            recovery_timeout: Seconds before attempting recovery
+            half_open_max_calls: Max calls allowed in half-open state
+        """
+        self.name = name
+        self.failure_threshold = failure_threshold
+        self.recovery_timeout = recovery_timeout
+        self.half_open_max_calls = half_open_max_calls
+
+        self._state = CircuitState.CLOSED
+        self._stats = CircuitStats()
+        self._lock = asyncio.Lock()
+
+    @property
+    def state(self) -> CircuitState:
+        """Get current circuit state (may trigger state transition)."""
+        self._check_state_transition()
+        return self._state
+
+    @property
+    def stats(self) -> CircuitStats:
+        """Get circuit statistics."""
+        return self._stats
+
+    def _check_state_transition(self) -> None:
+        """Check if state should transition based on time."""
+        if self._state == CircuitState.OPEN:
+            time_in_open = time.time() - self._stats.state_changed_at
+            if time_in_open >= self.recovery_timeout:
+                self._transition_to(CircuitState.HALF_OPEN)
+                logger.info(
+                    f"Circuit {self.name} transitioned to HALF_OPEN "
+                    f"after {time_in_open:.1f}s"
+                )
+
+    def _transition_to(self, new_state: CircuitState) -> None:
+        """Transition to a new state."""
+        old_state = self._state
+        self._state = new_state
+        self._stats.state_changed_at = time.time()
+
+        if new_state == CircuitState.HALF_OPEN:
+            self._stats.half_open_calls = 0
+        elif new_state == CircuitState.CLOSED:
+            self._stats.failures = 0
+
+        logger.debug(f"Circuit {self.name}: {old_state.value} -> {new_state.value}")
+
+    def is_available(self) -> bool:
+        """Check if circuit allows requests."""
+        state = self.state  # Triggers state check
+        if state == CircuitState.CLOSED:
+            return True
+        if state == CircuitState.HALF_OPEN:
+            return self._stats.half_open_calls < self.half_open_max_calls
+        return False
+
+    def time_until_recovery(self) -> int | None:
+        """Get seconds until circuit may recover (None if not open)."""
+        if self._state != CircuitState.OPEN:
+            return None
+        elapsed = time.time() - self._stats.state_changed_at
+        remaining = max(0, self.recovery_timeout - int(elapsed))
+        return remaining if remaining > 0 else 0
+
+    async def record_success(self) -> None:
+        """Record a successful call."""
+        async with self._lock:
+            self._stats.successes += 1
+            self._stats.last_success_time = time.time()
+
+            if self._state == CircuitState.HALF_OPEN:
+                # Success in half-open state closes the circuit
+                self._transition_to(CircuitState.CLOSED)
+                logger.info(f"Circuit {self.name} closed after successful recovery")
+
+    async def record_failure(self, error: Exception | None = None) -> None:  # noqa: ARG002
+        """
+        Record a failed call.
+
+        Args:
+            error: The exception that occurred
+        """
+        async with self._lock:
+            self._stats.failures += 1
+            self._stats.last_failure_time = time.time()
+
+            if self._state == CircuitState.HALF_OPEN:
+                # Failure in half-open state opens the circuit
+                self._transition_to(CircuitState.OPEN)
+                logger.warning(
+                    f"Circuit {self.name} reopened after failure in half-open state"
+                )
+            elif self._state == CircuitState.CLOSED:
+                if self._stats.failures >= self.failure_threshold:
+                    self._transition_to(CircuitState.OPEN)
+                    logger.warning(
+                        f"Circuit {self.name} opened after "
+                        f"{self._stats.failures} failures"
+                    )
+
+    async def execute(
+        self,
+        func: Callable[..., T],
+        *args: Any,
+        **kwargs: Any,
+    ) -> T:
+        """
+        Execute a function with circuit breaker protection.
+
+        Args:
+            func: Async function to execute
+            *args: Positional arguments
+            **kwargs: Keyword arguments
+
+        Returns:
+            Function result
+
+        Raises:
+            CircuitOpenError: If circuit is open
+        """
+        if not self.is_available():
+            raise CircuitOpenError(
+                provider=self.name,
+                recovery_time=self.time_until_recovery(),
+            )
+
+        async with self._lock:
+            if self._state == CircuitState.HALF_OPEN:
+                self._stats.half_open_calls += 1
+
+        try:
+            result = await func(*args, **kwargs)
+            await self.record_success()
+            return result
+        except Exception as e:
+            await self.record_failure(e)
+            raise
+
+    def reset(self) -> None:
+        """Reset circuit to closed state."""
+        self._state = CircuitState.CLOSED
+        self._stats = CircuitStats()
+        logger.info(f"Circuit {self.name} reset to CLOSED")
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert circuit state to dictionary."""
+        return {
+            "name": self.name,
+            "state": self._state.value,
+            "failures": self._stats.failures,
+            "successes": self._stats.successes,
+            "last_failure_time": self._stats.last_failure_time,
+            "last_success_time": self._stats.last_success_time,
+            "time_until_recovery": self.time_until_recovery(),
+            "is_available": self.is_available(),
+        }
+
+
+class CircuitBreakerRegistry:
+    """
+    Registry for managing multiple circuit breakers.
+
+    Provides centralized management of circuits for different providers/models.
+    """
+
+    def __init__(self, settings: Settings | None = None) -> None:
+        """
+        Initialize registry.
+
+        Args:
+            settings: Application settings (uses default if None)
+        """
+        self._settings = settings or get_settings()
+        self._circuits: dict[str, CircuitBreaker] = {}
+        self._lock = asyncio.Lock()
+
+    async def get_circuit(self, name: str) -> CircuitBreaker:
+        """
+        Get or create a circuit breaker.
+
+        Args:
+            name: Circuit identifier (e.g., provider name)
+
+        Returns:
+            CircuitBreaker instance
+        """
+        async with self._lock:
+            if name not in self._circuits:
+                self._circuits[name] = CircuitBreaker(
+                    name=name,
+                    failure_threshold=self._settings.circuit_failure_threshold,
+                    recovery_timeout=self._settings.circuit_recovery_timeout,
+                    half_open_max_calls=self._settings.circuit_half_open_max_calls,
+                )
+            return self._circuits[name]
+
+    def get_circuit_sync(self, name: str) -> CircuitBreaker:
+        """
+        Get or create a circuit breaker (sync version).
+
+        Args:
+            name: Circuit identifier
+
+        Returns:
+            CircuitBreaker instance
+        """
+        if name not in self._circuits:
+            self._circuits[name] = CircuitBreaker(
+                name=name,
+                failure_threshold=self._settings.circuit_failure_threshold,
+                recovery_timeout=self._settings.circuit_recovery_timeout,
+                half_open_max_calls=self._settings.circuit_half_open_max_calls,
+            )
+        return self._circuits[name]
+
+    async def is_available(self, name: str) -> bool:
+        """
+        Check if a circuit is available.
+
+        Args:
+            name: Circuit identifier
+
+        Returns:
+            True if circuit allows requests
+        """
+        circuit = await self.get_circuit(name)
+        return circuit.is_available()
+
+    async def record_success(self, name: str) -> None:
+        """Record success for a circuit."""
+        circuit = await self.get_circuit(name)
+        await circuit.record_success()
+
+    async def record_failure(self, name: str, error: Exception | None = None) -> None:
+        """Record failure for a circuit."""
+        circuit = await self.get_circuit(name)
+        await circuit.record_failure(error)
+
+    async def reset(self, name: str) -> None:
+        """Reset a specific circuit."""
+        async with self._lock:
+            if name in self._circuits:
+                self._circuits[name].reset()
+
+    async def reset_all(self) -> None:
+        """Reset all circuits."""
+        async with self._lock:
+            for circuit in self._circuits.values():
+                circuit.reset()
+
+    def get_all_states(self) -> dict[str, dict[str, Any]]:
+        """Get state of all circuits."""
+        return {name: circuit.to_dict() for name, circuit in self._circuits.items()}
+
+    def get_open_circuits(self) -> list[str]:
+        """Get list of circuits that are currently open."""
+        return [
+            name
+            for name, circuit in self._circuits.items()
+            if circuit.state == CircuitState.OPEN
+        ]
+
+    def get_available_circuits(self) -> list[str]:
+        """Get list of circuits that are available for requests."""
+        return [
+            name for name, circuit in self._circuits.items() if circuit.is_available()
+        ]
+
+
+# Global registry instance (lazy initialization)
+_registry: CircuitBreakerRegistry | None = None
+
+
+def get_circuit_registry() -> CircuitBreakerRegistry:
+    """Get the global circuit breaker registry."""
+    global _registry
+    if _registry is None:
+        _registry = CircuitBreakerRegistry()
+    return _registry
+
+
+def reset_circuit_registry() -> None:
+    """Reset the global registry (for testing)."""
+    global _registry
+    _registry = None
--- a/mcp-servers/llm-gateway/models.py
+++ b/mcp-servers/llm-gateway/models.py
@@ -0,0 +1,442 @@
+"""
+Data models for LLM Gateway MCP Server.
+
+Defines model groups, pricing, request/response structures.
+Per ADR-004: LLM Provider Abstraction.
+"""
+
+from dataclasses import dataclass, field
+from datetime import datetime
+from enum import Enum
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+
+class ModelGroup(str, Enum):
+    """Model groups for routing LLM requests."""
+
+    REASONING = "reasoning"  # Complex analysis, architecture decisions
+    CODE = "code"  # Code writing and refactoring
+    FAST = "fast"  # Quick tasks, simple queries
+    VISION = "vision"  # Multimodal image analysis
+    EMBEDDING = "embedding"  # Vector embeddings
+    COST_OPTIMIZED = "cost_optimized"  # High-volume, non-critical
+    SELF_HOSTED = "self_hosted"  # Privacy-sensitive, air-gapped
+
+    # Aliases for backward compatibility with ADR-004
+    HIGH_REASONING = "reasoning"
+    CODE_GENERATION = "code"
+    FAST_RESPONSE = "fast"
+
+
+class Provider(str, Enum):
+    """Supported LLM providers."""
+
+    ANTHROPIC = "anthropic"
+    OPENAI = "openai"
+    GOOGLE = "google"
+    ALIBABA = "alibaba"
+    DEEPSEEK = "deepseek"
+
+
+@dataclass
+class ModelConfig:
+    """Configuration for a specific model."""
+
+    name: str  # Model identifier (e.g., "claude-3-opus-20240229")
+    litellm_name: str  # LiteLLM model string (e.g., "anthropic/claude-3-opus-20240229")
+    provider: Provider
+    cost_per_1m_input: float  # USD per 1M input tokens
+    cost_per_1m_output: float  # USD per 1M output tokens
+    context_window: int  # Max context tokens
+    max_output_tokens: int  # Max output tokens
+    supports_vision: bool = False
+    supports_streaming: bool = True
+    supports_function_calling: bool = True
+
+
+# Model configurations per ADR-004
+MODEL_CONFIGS: dict[str, ModelConfig] = {
+    # Anthropic models
+    "claude-opus-4": ModelConfig(
+        name="claude-opus-4",
+        litellm_name="anthropic/claude-sonnet-4-20250514",  # Using sonnet-4 as opus-4 placeholder
+        provider=Provider.ANTHROPIC,
+        cost_per_1m_input=15.0,
+        cost_per_1m_output=75.0,
+        context_window=200000,
+        max_output_tokens=8192,
+        supports_vision=True,
+    ),
+    "claude-sonnet-4": ModelConfig(
+        name="claude-sonnet-4",
+        litellm_name="anthropic/claude-sonnet-4-20250514",
+        provider=Provider.ANTHROPIC,
+        cost_per_1m_input=3.0,
+        cost_per_1m_output=15.0,
+        context_window=200000,
+        max_output_tokens=8192,
+        supports_vision=True,
+    ),
+    "claude-haiku": ModelConfig(
+        name="claude-haiku",
+        litellm_name="anthropic/claude-3-5-haiku-20241022",
+        provider=Provider.ANTHROPIC,
+        cost_per_1m_input=1.0,
+        cost_per_1m_output=5.0,
+        context_window=200000,
+        max_output_tokens=8192,
+        supports_vision=True,
+    ),
+    # OpenAI models
+    "gpt-4.1": ModelConfig(
+        name="gpt-4.1",
+        litellm_name="openai/gpt-4.1",
+        provider=Provider.OPENAI,
+        cost_per_1m_input=2.0,
+        cost_per_1m_output=8.0,
+        context_window=1047576,
+        max_output_tokens=32768,
+        supports_vision=True,
+    ),
+    "gpt-4.1-mini": ModelConfig(
+        name="gpt-4.1-mini",
+        litellm_name="openai/gpt-4.1-mini",
+        provider=Provider.OPENAI,
+        cost_per_1m_input=0.4,
+        cost_per_1m_output=1.6,
+        context_window=1047576,
+        max_output_tokens=32768,
+        supports_vision=True,
+    ),
+    # Google models
+    "gemini-2.5-pro": ModelConfig(
+        name="gemini-2.5-pro",
+        litellm_name="gemini/gemini-2.5-pro",
+        provider=Provider.GOOGLE,
+        cost_per_1m_input=1.25,
+        cost_per_1m_output=10.0,
+        context_window=1048576,
+        max_output_tokens=65536,
+        supports_vision=True,
+    ),
+    "gemini-2.0-flash": ModelConfig(
+        name="gemini-2.0-flash",
+        litellm_name="gemini/gemini-2.0-flash",
+        provider=Provider.GOOGLE,
+        cost_per_1m_input=0.1,
+        cost_per_1m_output=0.4,
+        context_window=1048576,
+        max_output_tokens=8192,
+        supports_vision=True,
+    ),
+    # Alibaba models
+    "qwen-max": ModelConfig(
+        name="qwen-max",
+        litellm_name="alibaba/qwen-max",
+        provider=Provider.ALIBABA,
+        cost_per_1m_input=2.0,
+        cost_per_1m_output=6.0,
+        context_window=32768,
+        max_output_tokens=8192,
+        supports_vision=False,
+    ),
+    # DeepSeek models
+    "deepseek-coder": ModelConfig(
+        name="deepseek-coder",
+        litellm_name="deepseek/deepseek-coder",
+        provider=Provider.DEEPSEEK,
+        cost_per_1m_input=0.14,
+        cost_per_1m_output=0.28,
+        context_window=128000,
+        max_output_tokens=8192,
+        supports_vision=False,
+    ),
+    "deepseek-chat": ModelConfig(
+        name="deepseek-chat",
+        litellm_name="deepseek/deepseek-chat",
+        provider=Provider.DEEPSEEK,
+        cost_per_1m_input=0.14,
+        cost_per_1m_output=0.28,
+        context_window=128000,
+        max_output_tokens=8192,
+        supports_vision=False,
+    ),
+    # Embedding models
+    "text-embedding-3-large": ModelConfig(
+        name="text-embedding-3-large",
+        litellm_name="openai/text-embedding-3-large",
+        provider=Provider.OPENAI,
+        cost_per_1m_input=0.13,
+        cost_per_1m_output=0.0,
+        context_window=8191,
+        max_output_tokens=0,
+        supports_vision=False,
+        supports_streaming=False,
+        supports_function_calling=False,
+    ),
+    "voyage-3": ModelConfig(
+        name="voyage-3",
+        litellm_name="voyage/voyage-3",
+        provider=Provider.ANTHROPIC,  # Voyage via Anthropic partnership
+        cost_per_1m_input=0.06,
+        cost_per_1m_output=0.0,
+        context_window=32000,
+        max_output_tokens=0,
+        supports_vision=False,
+        supports_streaming=False,
+        supports_function_calling=False,
+    ),
+}
+
+
+@dataclass
+class ModelGroupConfig:
+    """Configuration for a model group with failover chain."""
+
+    primary: str  # Primary model name
+    fallbacks: list[str]  # Fallback model names in order
+    description: str
+
+    def get_all_models(self) -> list[str]:
+        """Get all models in priority order."""
+        return [self.primary, *self.fallbacks]
+
+
+# Model group configurations per ADR-004
+MODEL_GROUPS: dict[ModelGroup, ModelGroupConfig] = {
+    ModelGroup.REASONING: ModelGroupConfig(
+        primary="claude-opus-4",
+        fallbacks=["gpt-4.1", "gemini-2.5-pro", "qwen-max"],
+        description="Complex analysis, architecture decisions",
+    ),
+    ModelGroup.CODE: ModelGroupConfig(
+        primary="claude-sonnet-4",
+        fallbacks=["gpt-4.1", "deepseek-coder"],
+        description="Code writing and refactoring",
+    ),
+    ModelGroup.FAST: ModelGroupConfig(
+        primary="claude-haiku",
+        fallbacks=["gpt-4.1-mini", "gemini-2.0-flash"],
+        description="Quick tasks, simple queries",
+    ),
+    ModelGroup.VISION: ModelGroupConfig(
+        primary="claude-sonnet-4",
+        fallbacks=["gpt-4.1", "gemini-2.5-pro"],
+        description="Multimodal image analysis",
+    ),
+    ModelGroup.EMBEDDING: ModelGroupConfig(
+        primary="text-embedding-3-large",
+        fallbacks=["voyage-3"],
+        description="Vector embeddings",
+    ),
+    ModelGroup.COST_OPTIMIZED: ModelGroupConfig(
+        primary="qwen-max",
+        fallbacks=["deepseek-chat"],
+        description="High-volume, non-critical tasks",
+    ),
+    ModelGroup.SELF_HOSTED: ModelGroupConfig(
+        primary="deepseek-chat",
+        fallbacks=["qwen-max"],
+        description="Privacy-sensitive, air-gapped",
+    ),
+}
+
+
+# Agent type to model group mapping per ADR-004
+AGENT_TYPE_MODEL_PREFERENCES: dict[str, ModelGroup] = {
+    "product_owner": ModelGroup.REASONING,
+    "software_architect": ModelGroup.REASONING,
+    "software_engineer": ModelGroup.CODE,
+    "qa_engineer": ModelGroup.CODE,
+    "devops_engineer": ModelGroup.FAST,
+    "project_manager": ModelGroup.FAST,
+    "business_analyst": ModelGroup.REASONING,
+}
+
+
+class ChatMessage(BaseModel):
+    """A single chat message."""
+
+    role: str = Field(..., description="Message role: system, user, assistant, tool")
+    content: str | list[dict[str, Any]] = Field(..., description="Message content")
+    name: str | None = Field(default=None, description="Optional name for the message")
+    tool_calls: list[dict[str, Any]] | None = Field(
+        default=None, description="Tool calls if role is assistant"
+    )
+    tool_call_id: str | None = Field(
+        default=None, description="Tool call ID if role is tool"
+    )
+
+
+class CompletionRequest(BaseModel):
+    """Request for chat completion."""
+
+    project_id: str = Field(..., description="Project ID for cost attribution")
+    agent_id: str = Field(..., description="Agent ID making the request")
+    messages: list[ChatMessage] = Field(..., description="Chat messages")
+    model_group: ModelGroup = Field(
+        default=ModelGroup.REASONING, description="Model group for routing"
+    )
+    model_override: str | None = Field(
+        default=None, description="Specific model to use (bypasses routing)"
+    )
+    max_tokens: int = Field(default=4096, ge=1, le=32768, description="Max output tokens")
+    temperature: float = Field(
+        default=0.7, ge=0.0, le=2.0, description="Sampling temperature"
+    )
+    stream: bool = Field(default=False, description="Enable streaming response")
+    session_id: str | None = Field(
+        default=None, description="Session ID for conversation tracking"
+    )
+    metadata: dict[str, Any] = Field(
+        default_factory=dict, description="Additional metadata"
+    )
+
+
+class UsageStats(BaseModel):
+    """Token usage statistics."""
+
+    prompt_tokens: int = Field(default=0, description="Input tokens used")
+    completion_tokens: int = Field(default=0, description="Output tokens generated")
+    total_tokens: int = Field(default=0, description="Total tokens")
+    cost_usd: float = Field(default=0.0, description="Estimated cost in USD")
+
+    @classmethod
+    def from_response(
+        cls, prompt_tokens: int, completion_tokens: int, model_config: ModelConfig
+    ) -> "UsageStats":
+        """Create usage stats from token counts and model config."""
+        input_cost = (prompt_tokens / 1_000_000) * model_config.cost_per_1m_input
+        output_cost = (completion_tokens / 1_000_000) * model_config.cost_per_1m_output
+        return cls(
+            prompt_tokens=prompt_tokens,
+            completion_tokens=completion_tokens,
+            total_tokens=prompt_tokens + completion_tokens,
+            cost_usd=round(input_cost + output_cost, 6),
+        )
+
+
+class CompletionResponse(BaseModel):
+    """Response from chat completion."""
+
+    id: str = Field(..., description="Unique response ID")
+    model: str = Field(..., description="Model that generated the response")
+    provider: str = Field(..., description="Provider used")
+    content: str = Field(..., description="Generated content")
+    finish_reason: str = Field(
+        default="stop", description="Reason for completion finish"
+    )
+    usage: UsageStats = Field(default_factory=UsageStats, description="Token usage")
+    created_at: datetime = Field(
+        default_factory=datetime.utcnow, description="Response timestamp"
+    )
+    metadata: dict[str, Any] = Field(
+        default_factory=dict, description="Additional metadata"
+    )
+
+
+class StreamChunk(BaseModel):
+    """A chunk from a streaming response."""
+
+    id: str = Field(..., description="Chunk ID")
+    delta: str = Field(default="", description="Content delta")
+    finish_reason: str | None = Field(default=None, description="Finish reason if done")
+    usage: UsageStats | None = Field(
+        default=None, description="Usage stats (only on final chunk)"
+    )
+
+
+class EmbeddingRequest(BaseModel):
+    """Request for text embeddings."""
+
+    project_id: str = Field(..., description="Project ID for cost attribution")
+    agent_id: str = Field(..., description="Agent ID making the request")
+    texts: list[str] = Field(..., min_length=1, description="Texts to embed")
+    model: str = Field(
+        default="text-embedding-3-large", description="Embedding model"
+    )
+
+
+class EmbeddingResponse(BaseModel):
+    """Response from embedding generation."""
+
+    model: str = Field(..., description="Model used")
+    embeddings: list[list[float]] = Field(..., description="Embedding vectors")
+    usage: UsageStats = Field(default_factory=UsageStats, description="Token usage")
+
+
+@dataclass
+class CostRecord:
+    """A single cost record for tracking."""
+
+    project_id: str
+    agent_id: str
+    model: str
+    prompt_tokens: int
+    completion_tokens: int
+    cost_usd: float
+    timestamp: datetime = field(default_factory=datetime.utcnow)
+    session_id: str | None = None
+    request_id: str | None = None
+
+
+class UsageReport(BaseModel):
+    """Usage report for a project or agent."""
+
+    entity_id: str = Field(..., description="Project or agent ID")
+    entity_type: str = Field(..., description="'project' or 'agent'")
+    period: str = Field(..., description="Report period (hour, day, month)")
+    period_start: datetime = Field(..., description="Period start time")
+    period_end: datetime = Field(..., description="Period end time")
+    total_requests: int = Field(default=0, description="Total requests")
+    total_tokens: int = Field(default=0, description="Total tokens used")
+    total_cost_usd: float = Field(default=0.0, description="Total cost in USD")
+    by_model: dict[str, dict[str, Any]] = Field(
+        default_factory=dict, description="Breakdown by model"
+    )
+    by_agent: dict[str, dict[str, Any]] = Field(
+        default_factory=dict, description="Breakdown by agent (for project reports)"
+    )
+
+
+class ModelInfo(BaseModel):
+    """Information about an available model."""
+
+    name: str = Field(..., description="Model name")
+    provider: str = Field(..., description="Provider name")
+    cost_per_1m_input: float = Field(..., description="Input cost per 1M tokens")
+    cost_per_1m_output: float = Field(..., description="Output cost per 1M tokens")
+    context_window: int = Field(..., description="Max context tokens")
+    max_output_tokens: int = Field(..., description="Max output tokens")
+    supports_vision: bool = Field(..., description="Vision capability")
+    supports_streaming: bool = Field(..., description="Streaming capability")
+    supports_function_calling: bool = Field(..., description="Function calling")
+    available: bool = Field(default=True, description="Provider configured")
+
+    @classmethod
+    def from_config(cls, config: ModelConfig, available: bool = True) -> "ModelInfo":
+        """Create ModelInfo from ModelConfig."""
+        return cls(
+            name=config.name,
+            provider=config.provider.value,
+            cost_per_1m_input=config.cost_per_1m_input,
+            cost_per_1m_output=config.cost_per_1m_output,
+            context_window=config.context_window,
+            max_output_tokens=config.max_output_tokens,
+            supports_vision=config.supports_vision,
+            supports_streaming=config.supports_streaming,
+            supports_function_calling=config.supports_function_calling,
+            available=available,
+        )
+
+
+class ModelGroupInfo(BaseModel):
+    """Information about a model group."""
+
+    name: str = Field(..., description="Group name")
+    description: str = Field(..., description="Group description")
+    primary_model: str = Field(..., description="Primary model")
+    fallback_models: list[str] = Field(..., description="Fallback models")
+    available: bool = Field(default=True, description="At least one model available")
--- a/mcp-servers/llm-gateway/providers.py
+++ b/mcp-servers/llm-gateway/providers.py
@@ -0,0 +1,331 @@
+"""
+LiteLLM provider configuration for LLM Gateway.
+
+Configures the LiteLLM Router with model lists and failover chains.
+"""
+
+import logging
+import os
+from typing import Any
+
+import litellm
+from litellm import Router
+
+from config import Settings, get_settings
+from models import (
+    MODEL_CONFIGS,
+    MODEL_GROUPS,
+    ModelConfig,
+    ModelGroup,
+    Provider,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def configure_litellm(settings: Settings) -> None:
+    """
+    Configure LiteLLM global settings.
+
+    Args:
+        settings: Application settings
+    """
+    # Set API keys in environment (LiteLLM reads from env)
+    if settings.anthropic_api_key:
+        os.environ["ANTHROPIC_API_KEY"] = settings.anthropic_api_key
+    if settings.openai_api_key:
+        os.environ["OPENAI_API_KEY"] = settings.openai_api_key
+    if settings.google_api_key:
+        os.environ["GEMINI_API_KEY"] = settings.google_api_key
+    if settings.alibaba_api_key:
+        os.environ["DASHSCOPE_API_KEY"] = settings.alibaba_api_key
+    if settings.deepseek_api_key:
+        os.environ["DEEPSEEK_API_KEY"] = settings.deepseek_api_key
+
+    # Configure LiteLLM settings
+    litellm.drop_params = True  # Drop unsupported params instead of erroring
+    litellm.set_verbose = settings.debug
+
+    # Configure caching if enabled
+    if settings.litellm_cache_enabled:
+        litellm.cache = litellm.Cache(
+            type="redis",
+            host=_parse_redis_host(settings.redis_url),
+            port=_parse_redis_port(settings.redis_url),
+            ttl=settings.litellm_cache_ttl,
+        )
+
+
+def _parse_redis_host(redis_url: str) -> str:
+    """Extract host from Redis URL."""
+    # redis://host:port/db
+    url = redis_url.replace("redis://", "")
+    host_port = url.split("/")[0]
+    return host_port.split(":")[0]
+
+
+def _parse_redis_port(redis_url: str) -> int:
+    """Extract port from Redis URL."""
+    url = redis_url.replace("redis://", "")
+    host_port = url.split("/")[0]
+    parts = host_port.split(":")
+    return int(parts[1]) if len(parts) > 1 else 6379
+
+
+def _is_provider_available(provider: Provider, settings: Settings) -> bool:
+    """Check if a provider is available (API key configured)."""
+    provider_key_map = {
+        Provider.ANTHROPIC: settings.anthropic_api_key,
+        Provider.OPENAI: settings.openai_api_key,
+        Provider.GOOGLE: settings.google_api_key,
+        Provider.ALIBABA: settings.alibaba_api_key,
+        Provider.DEEPSEEK: settings.deepseek_api_key or settings.deepseek_base_url,
+    }
+    return bool(provider_key_map.get(provider))
+
+
+def _build_model_entry(
+    model_config: ModelConfig,
+    settings: Settings,
+) -> dict[str, Any] | None:
+    """
+    Build a model entry for LiteLLM Router.
+
+    Args:
+        model_config: Model configuration
+        settings: Application settings
+
+    Returns:
+        Model entry dict or None if provider unavailable
+    """
+    if not _is_provider_available(model_config.provider, settings):
+        logger.debug(
+            f"Skipping model {model_config.name}: "
+            f"{model_config.provider.value} provider not configured"
+        )
+        return None
+
+    entry: dict[str, Any] = {
+        "model_name": model_config.name,
+        "litellm_params": {
+            "model": model_config.litellm_name,
+            "timeout": settings.litellm_timeout,
+            "max_retries": settings.litellm_max_retries,
+        },
+    }
+
+    # Add custom base URL for DeepSeek self-hosted
+    if (
+        model_config.provider == Provider.DEEPSEEK
+        and settings.deepseek_base_url
+    ):
+        entry["litellm_params"]["api_base"] = settings.deepseek_base_url
+
+    return entry
+
+
+def build_model_list(settings: Settings | None = None) -> list[dict[str, Any]]:
+    """
+    Build the complete model list for LiteLLM Router.
+
+    Args:
+        settings: Application settings (uses default if None)
+
+    Returns:
+        List of model entries for Router
+    """
+    if settings is None:
+        settings = get_settings()
+
+    model_list: list[dict[str, Any]] = []
+
+    for model_config in MODEL_CONFIGS.values():
+        entry = _build_model_entry(model_config, settings)
+        if entry:
+            model_list.append(entry)
+
+    logger.info(f"Built model list with {len(model_list)} models")
+    return model_list
+
+
+def build_fallback_config(settings: Settings | None = None) -> dict[str, list[str]]:
+    """
+    Build fallback configuration based on model groups.
+
+    Args:
+        settings: Application settings (uses default if None)
+
+    Returns:
+        Dict mapping model names to their fallback chains
+    """
+    if settings is None:
+        settings = get_settings()
+
+    fallbacks: dict[str, list[str]] = {}
+
+    for _group, config in MODEL_GROUPS.items():
+        # Get available models in this group's chain
+        available_models = []
+        for model_name in config.get_all_models():
+            model_config = MODEL_CONFIGS.get(model_name)
+            if model_config and _is_provider_available(model_config.provider, settings):
+                available_models.append(model_name)
+
+        if len(available_models) > 1:
+            # First model falls back to remaining models
+            fallbacks[available_models[0]] = available_models[1:]
+
+    return fallbacks
+
+
+def get_available_models(settings: Settings | None = None) -> dict[str, ModelConfig]:
+    """
+    Get all available models (with configured providers).
+
+    Args:
+        settings: Application settings (uses default if None)
+
+    Returns:
+        Dict of available model configs
+    """
+    if settings is None:
+        settings = get_settings()
+
+    available: dict[str, ModelConfig] = {}
+
+    for name, config in MODEL_CONFIGS.items():
+        if _is_provider_available(config.provider, settings):
+            available[name] = config
+
+    return available
+
+
+def get_available_model_groups(
+    settings: Settings | None = None,
+) -> dict[ModelGroup, list[str]]:
+    """
+    Get available models for each model group.
+
+    Args:
+        settings: Application settings (uses default if None)
+
+    Returns:
+        Dict mapping model groups to available models
+    """
+    if settings is None:
+        settings = get_settings()
+
+    result: dict[ModelGroup, list[str]] = {}
+
+    for group, config in MODEL_GROUPS.items():
+        available_models = []
+        for model_name in config.get_all_models():
+            model_config = MODEL_CONFIGS.get(model_name)
+            if model_config and _is_provider_available(model_config.provider, settings):
+                available_models.append(model_name)
+        result[group] = available_models
+
+    return result
+
+
+class LLMProvider:
+    """
+    LLM Provider wrapper around LiteLLM Router.
+
+    Provides a high-level interface for LLM operations with
+    automatic failover and configuration management.
+    """
+
+    def __init__(self, settings: Settings | None = None) -> None:
+        """
+        Initialize LLM Provider.
+
+        Args:
+            settings: Application settings (uses default if None)
+        """
+        self._settings = settings or get_settings()
+        self._router: Router | None = None
+        self._initialized = False
+
+    def initialize(self) -> None:
+        """Initialize the provider and LiteLLM Router."""
+        if self._initialized:
+            return
+
+        # Configure LiteLLM global settings
+        configure_litellm(self._settings)
+
+        # Build model list
+        model_list = build_model_list(self._settings)
+
+        if not model_list:
+            logger.warning("No models available - no providers configured")
+            self._initialized = True
+            return
+
+        # Build fallback config
+        fallbacks = build_fallback_config(self._settings)
+
+        # Create Router
+        self._router = Router(
+            model_list=model_list,
+            fallbacks=list(fallbacks.items()) if fallbacks else None,
+            routing_strategy="latency-based-routing",
+            num_retries=self._settings.litellm_max_retries,
+            timeout=self._settings.litellm_timeout,
+            retry_after=5,  # Retry after 5 seconds
+            allowed_fails=2,  # Fail after 2 consecutive failures
+        )
+
+        self._initialized = True
+        logger.info(
+            f"LLM Provider initialized with {len(model_list)} models, "
+            f"{len(fallbacks)} fallback chains"
+        )
+
+    @property
+    def router(self) -> Router | None:
+        """Get the LiteLLM Router."""
+        if not self._initialized:
+            self.initialize()
+        return self._router
+
+    @property
+    def is_available(self) -> bool:
+        """Check if provider is available."""
+        if not self._initialized:
+            self.initialize()
+        return self._router is not None
+
+    def get_model_config(self, model_name: str) -> ModelConfig | None:
+        """Get configuration for a specific model."""
+        return MODEL_CONFIGS.get(model_name)
+
+    def get_available_models(self) -> dict[str, ModelConfig]:
+        """Get all available models."""
+        return get_available_models(self._settings)
+
+    def is_model_available(self, model_name: str) -> bool:
+        """Check if a specific model is available."""
+        model_config = MODEL_CONFIGS.get(model_name)
+        if not model_config:
+            return False
+        return _is_provider_available(model_config.provider, self._settings)
+
+
+# Global provider instance (lazy initialization)
+_provider: LLMProvider | None = None
+
+
+def get_provider() -> LLMProvider:
+    """Get the global LLM Provider instance."""
+    global _provider
+    if _provider is None:
+        _provider = LLMProvider()
+    return _provider
+
+
+def reset_provider() -> None:
+    """Reset the global provider (for testing)."""
+    global _provider
+    _provider = None
--- a/mcp-servers/llm-gateway/pyproject.toml
+++ b/mcp-servers/llm-gateway/pyproject.toml
@@ -4,20 +4,96 @@ version = "0.1.0"
 description = "Syndarix LLM Gateway MCP Server - Unified LLM access with failover and cost tracking"
 requires-python = ">=3.12"
 dependencies = [
-    "fastmcp>=0.1.0",
+    "fastmcp>=2.0.0",
    "litellm>=1.50.0",
    "redis>=5.0.0",
    "pydantic>=2.0.0",
    "pydantic-settings>=2.0.0",
+    "tiktoken>=0.7.0",
+    "httpx>=0.27.0",
+    "uvicorn>=0.30.0",
+    "fastapi>=0.115.0",
 ]

 [project.optional-dependencies]
 dev = [
    "pytest>=8.0.0",
-    "pytest-asyncio>=0.23.0",
+    "pytest-asyncio>=0.24.0",
+    "pytest-cov>=5.0.0",
+    "respx>=0.21.0",
+    "fakeredis>=2.25.0",
    "ruff>=0.8.0",
+    "mypy>=1.11.0",
 ]

+[project.scripts]
+llm-gateway = "server:main"
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["."]
+exclude = ["tests/", "*.md", "Dockerfile"]
+
+[tool.hatch.build.targets.sdist]
+include = ["*.py", "pyproject.toml"]
+
 [tool.ruff]
 target-version = "py312"
 line-length = 88
+
+[tool.ruff.lint]
+select = [
+    "E",      # pycodestyle errors
+    "W",      # pycodestyle warnings
+    "F",      # pyflakes
+    "I",      # isort
+    "B",      # flake8-bugbear
+    "C4",     # flake8-comprehensions
+    "UP",     # pyupgrade
+    "ARG",    # flake8-unused-arguments
+    "SIM",    # flake8-simplify
+]
+ignore = [
+    "E501",   # line too long (handled by formatter)
+    "B008",   # do not perform function calls in argument defaults
+    "B904",   # raise from in except (too noisy)
+]
+
+[tool.ruff.lint.isort]
+known-first-party = ["config", "models", "exceptions", "providers", "failover", "routing", "cost_tracking", "streaming"]
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+asyncio_default_fixture_loop_scope = "function"
+testpaths = ["tests"]
+addopts = "-v --tb=short"
+filterwarnings = [
+    "ignore::DeprecationWarning",
+]
+
+[tool.coverage.run]
+source = ["."]
+omit = ["tests/*", "conftest.py"]
+branch = true
+
+[tool.coverage.report]
+exclude_lines = [
+    "pragma: no cover",
+    "def __repr__",
+    "raise NotImplementedError",
+    "if TYPE_CHECKING:",
+    "if __name__ == .__main__.:",
+]
+fail_under = 90
+show_missing = true
+
+[tool.mypy]
+python_version = "3.12"
+strict = true
+warn_return_any = true
+warn_unused_ignores = true
+disallow_untyped_defs = true
+plugins = ["pydantic.mypy"]
--- a/mcp-servers/llm-gateway/routing.py
+++ b/mcp-servers/llm-gateway/routing.py
@@ -0,0 +1,319 @@
+"""
+Model routing for LLM Gateway.
+
+Handles model selection based on:
+- Model group configuration
+- Circuit breaker availability
+- Agent type preferences
+"""
+
+import logging
+from typing import Any
+
+from config import Settings, get_settings
+from exceptions import (
+    AllProvidersFailedError,
+    InvalidModelError,
+    InvalidModelGroupError,
+    ModelNotAvailableError,
+)
+from failover import CircuitBreakerRegistry, get_circuit_registry
+from models import (
+    AGENT_TYPE_MODEL_PREFERENCES,
+    MODEL_CONFIGS,
+    MODEL_GROUPS,
+    ModelConfig,
+    ModelGroup,
+)
+from providers import get_available_models
+
+logger = logging.getLogger(__name__)
+
+
+class ModelRouter:
+    """
+    Routes requests to appropriate models based on configuration.
+
+    Considers:
+    - Model group preferences
+    - Circuit breaker states
+    - Agent type defaults
+    - Provider availability
+    """
+
+    def __init__(
+        self,
+        settings: Settings | None = None,
+        circuit_registry: CircuitBreakerRegistry | None = None,
+    ) -> None:
+        """
+        Initialize model router.
+
+        Args:
+            settings: Application settings
+            circuit_registry: Circuit breaker registry
+        """
+        self._settings = settings or get_settings()
+        self._circuit_registry = circuit_registry or get_circuit_registry()
+
+    def parse_model_group(self, group_str: str) -> ModelGroup:
+        """
+        Parse model group from string.
+
+        Args:
+            group_str: Group name string
+
+        Returns:
+            ModelGroup enum value
+
+        Raises:
+            InvalidModelGroupError: If group is unknown
+        """
+        # Handle aliases
+        aliases = {
+            "high-reasoning": ModelGroup.REASONING,
+            "high_reasoning": ModelGroup.REASONING,
+            "code-generation": ModelGroup.CODE,
+            "code_generation": ModelGroup.CODE,
+            "fast-response": ModelGroup.FAST,
+            "fast_response": ModelGroup.FAST,
+        }
+
+        # Try direct enum value
+        try:
+            return ModelGroup(group_str.lower())
+        except ValueError:
+            pass
+
+        # Try aliases
+        if group_str.lower() in aliases:
+            return aliases[group_str.lower()]
+
+        # Unknown group
+        available = [g.value for g in ModelGroup]
+        raise InvalidModelGroupError(
+            model_group=group_str,
+            available_groups=available,
+        )
+
+    def get_model_config(self, model_name: str) -> ModelConfig:
+        """
+        Get configuration for a specific model.
+
+        Args:
+            model_name: Model name
+
+        Returns:
+            Model configuration
+
+        Raises:
+            InvalidModelError: If model is unknown
+        """
+        config = MODEL_CONFIGS.get(model_name)
+        if not config:
+            raise InvalidModelError(
+                model=model_name,
+                reason="Unknown model",
+            )
+        return config
+
+    def get_preferred_group_for_agent(self, agent_type: str) -> ModelGroup:
+        """
+        Get preferred model group for an agent type.
+
+        Args:
+            agent_type: Agent type identifier
+
+        Returns:
+            Preferred ModelGroup
+        """
+        return AGENT_TYPE_MODEL_PREFERENCES.get(
+            agent_type.lower(),
+            ModelGroup.REASONING,  # Default to reasoning
+        )
+
+    async def select_model(
+        self,
+        model_group: ModelGroup | str,
+        model_override: str | None = None,
+        agent_type: str | None = None,
+    ) -> tuple[str, ModelConfig]:
+        """
+        Select the best available model.
+
+        Args:
+            model_group: Desired model group
+            model_override: Specific model to use (bypasses group routing)
+            agent_type: Agent type for preference lookup
+
+        Returns:
+            Tuple of (model_name, model_config)
+
+        Raises:
+            InvalidModelError: If override model is invalid
+            InvalidModelGroupError: If group is invalid
+            AllProvidersFailedError: If no models are available
+        """
+        # Handle model override
+        if model_override:
+            config = MODEL_CONFIGS.get(model_override)
+            if not config:
+                raise InvalidModelError(
+                    model=model_override,
+                    reason="Unknown model",
+                )
+            # Check if model's provider is available (using router's settings)
+            available_models = get_available_models(self._settings)
+            if model_override not in available_models:
+                raise ModelNotAvailableError(
+                    model=model_override,
+                    provider=config.provider.value,
+                )
+            # Check circuit breaker
+            circuit = self._circuit_registry.get_circuit_sync(config.provider.value)
+            if not circuit.is_available():
+                raise ModelNotAvailableError(
+                    model=model_override,
+                    provider=f"{config.provider.value} (circuit open)",
+                )
+            return model_override, config
+
+        # Parse model group if string
+        if isinstance(model_group, str):
+            model_group = self.parse_model_group(model_group)
+
+        # Get agent type preference if no explicit group
+        if agent_type:
+            preferred = self.get_preferred_group_for_agent(agent_type)
+            logger.debug(
+                f"Agent type {agent_type} prefers {preferred.value}, "
+                f"requested {model_group.value}"
+            )
+
+        # Get group configuration
+        group_config = MODEL_GROUPS.get(model_group)
+        if not group_config:
+            raise InvalidModelGroupError(
+                model_group=model_group.value,
+                available_groups=[g.value for g in ModelGroup],
+            )
+
+        # Get available models
+        available_models = get_available_models(self._settings)
+
+        # Try models in priority order
+        errors: list[dict[str, Any]] = []
+        attempted: list[str] = []
+
+        for model_name in group_config.get_all_models():
+            attempted.append(model_name)
+
+            # Check if model provider is configured
+            config = MODEL_CONFIGS.get(model_name)
+            if not config:
+                errors.append({"model": model_name, "error": "Unknown model"})
+                continue
+
+            if model_name not in available_models:
+                errors.append({
+                    "model": model_name,
+                    "error": f"Provider {config.provider.value} not configured",
+                })
+                continue
+
+            # Check circuit breaker
+            circuit = self._circuit_registry.get_circuit_sync(config.provider.value)
+            if not circuit.is_available():
+                errors.append({
+                    "model": model_name,
+                    "error": f"Circuit open for {config.provider.value}",
+                })
+                continue
+
+            # Model is available
+            logger.debug(
+                f"Selected model {model_name} for group {model_group.value}"
+            )
+            return model_name, config
+
+        # No models available
+        raise AllProvidersFailedError(
+            model_group=model_group.value,
+            attempted_models=attempted,
+            errors=errors,
+        )
+
+    async def get_available_models_for_group(
+        self,
+        model_group: ModelGroup | str,
+    ) -> list[tuple[str, ModelConfig, bool]]:
+        """
+        Get all models for a group with availability status.
+
+        Args:
+            model_group: Model group
+
+        Returns:
+            List of (model_name, config, is_available) tuples
+        """
+        # Parse model group if string
+        if isinstance(model_group, str):
+            model_group = self.parse_model_group(model_group)
+
+        group_config = MODEL_GROUPS.get(model_group)
+        if not group_config:
+            return []
+
+        available_models = get_available_models(self._settings)
+        result: list[tuple[str, ModelConfig, bool]] = []
+
+        for model_name in group_config.get_all_models():
+            config = MODEL_CONFIGS.get(model_name)
+            if not config:
+                continue
+
+            is_available = model_name in available_models
+            if is_available:
+                # Also check circuit breaker
+                circuit = self._circuit_registry.get_circuit_sync(config.provider.value)
+                is_available = circuit.is_available()
+
+            result.append((model_name, config, is_available))
+
+        return result
+
+    def get_all_model_groups(self) -> dict[str, dict[str, Any]]:
+        """
+        Get information about all model groups.
+
+        Returns:
+            Dict of group info
+        """
+        result: dict[str, dict[str, Any]] = {}
+
+        for group, config in MODEL_GROUPS.items():
+            result[group.value] = {
+                "description": config.description,
+                "primary": config.primary,
+                "fallbacks": config.fallbacks,
+            }
+
+        return result
+
+
+# Global router instance (lazy initialization)
+_router: ModelRouter | None = None
+
+
+def get_model_router() -> ModelRouter:
+    """Get the global model router instance."""
+    global _router
+    if _router is None:
+        _router = ModelRouter()
+    return _router
+
+
+def reset_model_router() -> None:
+    """Reset the global router (for testing)."""
+    global _router
+    _router = None
--- a/mcp-servers/llm-gateway/server.py
+++ b/mcp-servers/llm-gateway/server.py
@@ -4,36 +4,555 @@ Syndarix LLM Gateway MCP Server.
 Provides unified LLM access with:
 - Multi-provider support (Claude, GPT, Gemini, Qwen, DeepSeek)
 - Automatic failover chains
- Cost tracking via LiteLLM callbacks
- Model group routing (high-reasoning, code-generation, fast-response, cost-optimized)
+- Cost tracking via Redis
+- Model group routing (reasoning, code, fast, vision, embedding)
+- Circuit breaker protection

 Per ADR-004: LLM Provider Abstraction.
 """

-import os
+import logging
+import uuid
+from contextlib import asynccontextmanager
+from typing import Any

+import tiktoken
+from fastapi import FastAPI
 from fastmcp import FastMCP

-# Create MCP server
-mcp = FastMCP(
-    "syndarix-llm-gateway",
-    description="Unified LLM access with failover and cost tracking",
+from config import get_settings
+from cost_tracking import calculate_cost, get_cost_tracker
+from exceptions import (
+    AllProvidersFailedError,
+    CircuitOpenError,
+    CostLimitExceededError,
+    InvalidModelError,
+    InvalidModelGroupError,
+    LLMGatewayError,
+    ModelNotAvailableError,
+)
+from failover import get_circuit_registry
+from models import (
+    MODEL_CONFIGS,
+    MODEL_GROUPS,
+    ModelGroup,
+)
+from providers import get_provider
+from routing import get_model_router
+
+# Configure logging
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+)
+logger = logging.getLogger(__name__)
+
+
+# Create FastMCP server
+mcp = FastMCP("syndarix-llm-gateway")
+
+
+@asynccontextmanager
+async def lifespan(_app: FastAPI):
+    """Application lifespan handler."""
+    settings = get_settings()
+    logger.info(f"Starting LLM Gateway on {settings.host}:{settings.port}")
+
+    # Initialize provider
+    provider = get_provider()
+    provider.initialize()
+
+    yield
+
+    # Cleanup
+    from cost_tracking import close_cost_tracker
+    await close_cost_tracker()
+    logger.info("LLM Gateway shutdown complete")
+
+
+# Create FastAPI app that wraps FastMCP
+app = FastAPI(
+    title="Syndarix LLM Gateway",
+    description="MCP Server for unified LLM access",
+    version="0.1.0",
+    lifespan=lifespan,
 )

-# Configuration
-REDIS_URL = os.getenv("REDIS_URL", "redis://localhost:6379/0")
-DATABASE_URL = os.getenv("DATABASE_URL")
+
+# Health endpoint
+@app.get("/health")
+async def health_check() -> dict[str, Any]:
+    """Health check endpoint."""
+    settings = get_settings()
+    provider = get_provider()
+
+    return {
+        "status": "healthy",
+        "service": "llm-gateway",
+        "providers_configured": settings.get_available_providers(),
+        "provider_available": provider.is_available,
+    }
+
+
+# Tool discovery endpoint (for MCP client compatibility)
+@app.get("/mcp/tools")
+async def list_tools() -> dict[str, Any]:
+    """List available MCP tools."""
+    return {
+        "tools": [
+            {
+                "name": "chat_completion",
+                "description": "Generate a chat completion using the specified model group",
+                "inputSchema": {
+                    "type": "object",
+                    "properties": {
+                        "project_id": {"type": "string", "description": "Project ID"},
+                        "agent_id": {"type": "string", "description": "Agent ID"},
+                        "messages": {
+                            "type": "array",
+                            "items": {
+                                "type": "object",
+                                "properties": {
+                                    "role": {"type": "string"},
+                                    "content": {"type": "string"},
+                                },
+                                "required": ["role", "content"],
+                            },
+                        },
+                        "model_group": {
+                            "type": "string",
+                            "enum": [g.value for g in ModelGroup],
+                            "default": "reasoning",
+                        },
+                        "max_tokens": {"type": "integer", "default": 4096},
+                        "temperature": {"type": "number", "default": 0.7},
+                        "stream": {"type": "boolean", "default": False},
+                    },
+                    "required": ["project_id", "agent_id", "messages"],
+                },
+            },
+            {
+                "name": "list_models",
+                "description": "List available models and model groups",
+                "inputSchema": {
+                    "type": "object",
+                    "properties": {
+                        "project_id": {"type": "string"},
+                        "agent_id": {"type": "string"},
+                        "model_group": {"type": "string"},
+                    },
+                    "required": ["project_id", "agent_id"],
+                },
+            },
+            {
+                "name": "get_usage",
+                "description": "Get usage statistics for a project or agent",
+                "inputSchema": {
+                    "type": "object",
+                    "properties": {
+                        "project_id": {"type": "string"},
+                        "agent_id": {"type": "string"},
+                        "period": {
+                            "type": "string",
+                            "enum": ["hour", "day", "month"],
+                            "default": "day",
+                        },
+                    },
+                    "required": ["project_id", "agent_id"],
+                },
+            },
+            {
+                "name": "count_tokens",
+                "description": "Count tokens in text",
+                "inputSchema": {
+                    "type": "object",
+                    "properties": {
+                        "project_id": {"type": "string"},
+                        "agent_id": {"type": "string"},
+                        "text": {"type": "string"},
+                        "model": {"type": "string"},
+                    },
+                    "required": ["project_id", "agent_id", "text"],
+                },
+            },
+        ]
+    }
+
+
+# JSON-RPC endpoint (for MCP client compatibility)
+@app.post("/mcp")
+async def jsonrpc_handler(request: dict[str, Any]) -> dict[str, Any]:
+    """Handle JSON-RPC 2.0 requests for MCP tools."""
+    # Validate JSON-RPC structure
+    if request.get("jsonrpc") != "2.0":
+        return {
+            "jsonrpc": "2.0",
+            "error": {"code": -32600, "message": "Invalid JSON-RPC version"},
+            "id": request.get("id"),
+        }
+
+    method = request.get("method")
+    params = request.get("params", {})
+    request_id = request.get("id")
+
+    # Handle tool calls
+    if method == "tools/call":
+        tool_name = params.get("name")
+        tool_args = params.get("arguments", {})
+
+        try:
+            if tool_name == "chat_completion":
+                result = await _impl_chat_completion(**tool_args)
+            elif tool_name == "list_models":
+                result = await _impl_list_models(**tool_args)
+            elif tool_name == "get_usage":
+                result = await _impl_get_usage(**tool_args)
+            elif tool_name == "count_tokens":
+                result = await _impl_count_tokens(**tool_args)
+            else:
+                return {
+                    "jsonrpc": "2.0",
+                    "error": {"code": -32601, "message": f"Unknown tool: {tool_name}"},
+                    "id": request_id,
+                }
+
+            return {
+                "jsonrpc": "2.0",
+                "result": {"content": [{"type": "text", "text": str(result)}]},
+                "id": request_id,
+            }
+
+        except LLMGatewayError as e:
+            return {
+                "jsonrpc": "2.0",
+                "error": {"code": -32000, "message": str(e), "data": e.to_dict()},
+                "id": request_id,
+            }
+        except Exception as e:
+            logger.exception(f"Error executing tool {tool_name}")
+            return {
+                "jsonrpc": "2.0",
+                "error": {"code": -32603, "message": str(e)},
+                "id": request_id,
+            }
+
+    # Handle tool listing
+    elif method == "tools/list":
+        tools_response = await list_tools()
+        return {
+            "jsonrpc": "2.0",
+            "result": tools_response,
+            "id": request_id,
+        }
+
+    else:
+        return {
+            "jsonrpc": "2.0",
+            "error": {"code": -32601, "message": f"Unknown method: {method}"},
+            "id": request_id,
+        }
+
+
+# ============================================================================
+# Core Implementation Functions
+# ============================================================================
+
+
+async def _impl_chat_completion(
+    project_id: str,
+    agent_id: str,
+    messages: list[dict[str, Any]],
+    model_group: str = "reasoning",
+    model_override: str | None = None,
+    max_tokens: int = 4096,
+    temperature: float = 0.7,
+    stream: bool = False,
+    session_id: str | None = None,
+) -> dict[str, Any]:
+    """Core implementation for chat completion."""
+    settings = get_settings()
+    router = get_model_router()
+    tracker = get_cost_tracker()
+    circuit_registry = get_circuit_registry()
+
+    # Check budget before making request
+    if settings.cost_tracking_enabled:
+        within_budget, current_cost, limit = await tracker.check_budget(project_id)
+        if not within_budget:
+            raise CostLimitExceededError(
+                entity_type="project",
+                entity_id=project_id,
+                current_cost=current_cost,
+                limit=limit,
+            )
+
+    # Select model
+    try:
+        model_name, model_config = await router.select_model(
+            model_group=model_group,
+            model_override=model_override,
+        )
+    except (InvalidModelGroupError, InvalidModelError, AllProvidersFailedError):
+        raise
+    except ModelNotAvailableError:
+        raise
+
+    # Get provider
+    provider = get_provider()
+    if not provider.router:
+        raise AllProvidersFailedError(
+            model_group=model_group,
+            attempted_models=[model_name],
+            errors=[{"error": "No providers configured"}],
+        )
+
+    # Generate request ID
+    request_id = str(uuid.uuid4())
+
+    try:
+        # Get circuit breaker for this provider
+        circuit = circuit_registry.get_circuit_sync(model_config.provider.value)
+
+        # Make completion request
+        if stream:
+            # Return streaming response info
+            # Actual streaming would be handled by a separate endpoint
+            return {
+                "status": "streaming_not_supported_via_tool",
+                "message": "Use /stream endpoint for streaming responses",
+                "request_id": request_id,
+            }
+
+        # Non-streaming completion
+        response = await provider.router.acompletion(
+            model=model_name,
+            messages=messages,
+            max_tokens=max_tokens,
+            temperature=temperature,
+        )
+
+        # Record success
+        await circuit.record_success()
+
+        # Extract response data
+        content = response.choices[0].message.content or ""
+        finish_reason = response.choices[0].finish_reason or "stop"
+
+        # Get usage stats
+        prompt_tokens = response.usage.prompt_tokens if response.usage else 0
+        completion_tokens = response.usage.completion_tokens if response.usage else 0
+
+        # Calculate cost
+        cost_usd = calculate_cost(model_name, prompt_tokens, completion_tokens)
+
+        # Record usage
+        await tracker.record_usage(
+            project_id=project_id,
+            agent_id=agent_id,
+            model=model_name,
+            prompt_tokens=prompt_tokens,
+            completion_tokens=completion_tokens,
+            cost_usd=cost_usd,
+            session_id=session_id,
+            request_id=request_id,
+        )
+
+        return {
+            "id": request_id,
+            "model": model_name,
+            "provider": model_config.provider.value,
+            "content": content,
+            "finish_reason": finish_reason,
+            "usage": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+                "cost_usd": cost_usd,
+            },
+        }
+
+    except CircuitOpenError:
+        raise
+    except Exception as e:
+        # Record failure
+        circuit = circuit_registry.get_circuit_sync(model_config.provider.value)
+        await circuit.record_failure(e)
+
+        logger.error(f"Completion failed: {e}")
+        raise AllProvidersFailedError(
+            model_group=model_group,
+            attempted_models=[model_name],
+            errors=[{"model": model_name, "error": str(e)}],
+        )
+
+
+async def _impl_list_models(
+    project_id: str,
+    agent_id: str,
+    model_group: str | None = None,
+) -> dict[str, Any]:
+    """Core implementation for list_models."""
+    settings = get_settings()
+    provider = get_provider()
+    router = get_model_router()
+
+    # Get available providers
+    available_providers = settings.get_available_providers()
+
+    result: dict[str, Any] = {
+        "project_id": project_id,
+        "agent_id": agent_id,
+        "available_providers": available_providers,
+    }
+
+    if model_group:
+        # List models for specific group
+        try:
+            parsed_group = router.parse_model_group(model_group)
+            models = await router.get_available_models_for_group(parsed_group)
+
+            result["model_group"] = model_group
+            result["models"] = [
+                {
+                    "name": name,
+                    "provider": config.provider.value,
+                    "available": available,
+                    "cost_per_1m_input": config.cost_per_1m_input,
+                    "cost_per_1m_output": config.cost_per_1m_output,
+                }
+                for name, config, available in models
+            ]
+        except InvalidModelGroupError as e:
+            result["error"] = e.to_dict()
+    else:
+        # List all model groups
+        groups: dict[str, Any] = {}
+        for group in ModelGroup:
+            group_config = MODEL_GROUPS.get(group)
+            if group_config:
+                models = await router.get_available_models_for_group(group)
+                available_count = sum(1 for _, _, avail in models if avail)
+                groups[group.value] = {
+                    "description": group_config.description,
+                    "primary": group_config.primary,
+                    "fallbacks": group_config.fallbacks,
+                    "available_models": available_count,
+                    "total_models": len(models),
+                }
+        result["model_groups"] = groups
+
+        # List all models
+        all_models: list[dict[str, Any]] = []
+        available_models = provider.get_available_models()
+        for name, config in MODEL_CONFIGS.items():
+            all_models.append({
+                "name": name,
+                "provider": config.provider.value,
+                "available": name in available_models,
+                "cost_per_1m_input": config.cost_per_1m_input,
+                "cost_per_1m_output": config.cost_per_1m_output,
+                "context_window": config.context_window,
+                "max_output_tokens": config.max_output_tokens,
+                "supports_vision": config.supports_vision,
+                "supports_streaming": config.supports_streaming,
+            })
+        result["models"] = all_models
+
+    return result
+
+
+async def _impl_get_usage(
+    project_id: str,
+    agent_id: str,
+    period: str = "day",
+) -> dict[str, Any]:
+    """Core implementation for get_usage."""
+    tracker = get_cost_tracker()
+
+    # Get project usage
+    project_report = await tracker.get_project_usage(project_id, period=period)
+
+    # Get agent usage
+    agent_report = await tracker.get_agent_usage(agent_id, period=period)
+
+    return {
+        "project_id": project_id,
+        "agent_id": agent_id,
+        "period": period,
+        "project_usage": {
+            "total_requests": project_report.total_requests,
+            "total_tokens": project_report.total_tokens,
+            "total_cost_usd": project_report.total_cost_usd,
+            "by_model": project_report.by_model,
+            "period_start": project_report.period_start.isoformat(),
+            "period_end": project_report.period_end.isoformat(),
+        },
+        "agent_usage": {
+            "total_requests": agent_report.total_requests,
+            "total_tokens": agent_report.total_tokens,
+            "total_cost_usd": agent_report.total_cost_usd,
+            "by_model": agent_report.by_model,
+            "period_start": agent_report.period_start.isoformat(),
+            "period_end": agent_report.period_end.isoformat(),
+        },
+    }
+
+
+async def _impl_count_tokens(
+    project_id: str,
+    agent_id: str,
+    text: str,
+    model: str | None = None,
+) -> dict[str, Any]:
+    """Core implementation for count_tokens."""
+    # Use tiktoken for token counting
+    # Default to cl100k_base (used by GPT-4, Claude, etc.)
+    try:
+        if model and model.startswith("gpt"):
+            encoding = tiktoken.encoding_for_model(model)
+        else:
+            encoding = tiktoken.get_encoding("cl100k_base")
+
+        token_count = len(encoding.encode(text))
+    except Exception as e:
+        logger.warning(f"Token counting failed: {e}, using estimate")
+        # Fallback: rough estimate of ~4 chars per token
+        token_count = len(text) // 4
+
+    # Estimate costs for different models
+    cost_estimates: dict[str, float] = {}
+    for model_name, config in MODEL_CONFIGS.items():
+        if config.cost_per_1m_input > 0:
+            cost = (token_count / 1_000_000) * config.cost_per_1m_input
+            cost_estimates[model_name] = round(cost, 6)
+
+    return {
+        "project_id": project_id,
+        "agent_id": agent_id,
+        "token_count": token_count,
+        "text_length": len(text),
+        "encoding": "cl100k_base",
+        "cost_estimates": cost_estimates,
+    }
+
+
+# ============================================================================
+# MCP Tools (wrappers around core implementations)
+# ============================================================================


@mcp.tool()
 async def chat_completion(
    project_id: str,
    agent_id: str,
-    messages: list[dict],
-    model_group: str = "high-reasoning",
+    messages: list[dict[str, Any]],
+    model_group: str = "reasoning",
+    model_override: str | None = None,
    max_tokens: int = 4096,
    temperature: float = 0.7,
-) -> dict:
+    stream: bool = False,
+    session_id: str | None = None,
+) -> dict[str, Any]:
    """
    Generate a chat completion using the specified model group.

@@ -41,108 +560,116 @@ async def chat_completion(
        project_id: UUID of the project (required for cost attribution)
        agent_id: UUID of the agent instance making the request
        messages: List of message dicts with 'role' and 'content'
-        model_group: Model routing group (high-reasoning, code-generation, fast-response, cost-optimized, self-hosted)
+        model_group: Model routing group (reasoning, code, fast, vision, embedding)
+        model_override: Specific model to use (bypasses group routing)
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature (0.0-2.0)
+        stream: Enable streaming response
+        session_id: Optional session ID for tracking

    Returns:
        Completion response with content and usage statistics
    """
-    # TODO: Implement with LiteLLM
-    # 1. Map model_group to primary model + fallbacks
-    # 2. Check project budget before making request
-    # 3. Make completion request with failover
-    # 4. Log usage via callback
-    # 5. Return response
-    return {
-        "status": "not_implemented",
-        "project_id": project_id,
-        "agent_id": agent_id,
-        "model_group": model_group,
-    }
+    return await _impl_chat_completion(
+        project_id=project_id,
+        agent_id=agent_id,
+        messages=messages,
+        model_group=model_group,
+        model_override=model_override,
+        max_tokens=max_tokens,
+        temperature=temperature,
+        stream=stream,
+        session_id=session_id,
+    )


@mcp.tool()
-async def get_embeddings(
+async def list_models(
    project_id: str,
-    texts: list[str],
-    model: str = "text-embedding-3-small",
-) -> dict:
+    agent_id: str,
+    model_group: str | None = None,
+) -> dict[str, Any]:
    """
-    Generate embeddings for the given texts.
-
-    Args:
-        project_id: UUID of the project (required for cost attribution)
-        texts: List of texts to embed
-        model: Embedding model to use
-
-    Returns:
-        List of embedding vectors
-    """
-    # TODO: Implement with LiteLLM embeddings
-    return {
-        "status": "not_implemented",
-        "project_id": project_id,
-        "text_count": len(texts),
-    }
-
-
-@mcp.tool()
-async def get_budget_status(project_id: str) -> dict:
-    """
-    Get current budget status for a project.
+    List available models and model groups.

    Args:
        project_id: UUID of the project
+        agent_id: UUID of the agent instance
+        model_group: Optional specific group to list

    Returns:
-        Budget status with usage, limits, and percentage
+        Dictionary of available models and groups
    """
-    # TODO: Implement budget check from Redis
-    return {
-        "status": "not_implemented",
-        "project_id": project_id,
-    }
+    return await _impl_list_models(
+        project_id=project_id,
+        agent_id=agent_id,
+        model_group=model_group,
+    )


@mcp.tool()
-async def list_available_models() -> dict:
+async def get_usage(
+    project_id: str,
+    agent_id: str,
+    period: str = "day",
+) -> dict[str, Any]:
    """
-    List all available models and their capabilities.
+    Get usage statistics for a project or agent.
+
+    Args:
+        project_id: UUID of the project
+        agent_id: UUID of the agent
+        period: Time period (hour, day, month)

    Returns:
-        Dictionary of model groups and available models
+        Usage statistics including tokens and costs
    """
-    return {
-        "model_groups": {
-            "high-reasoning": {
-                "primary": "claude-opus-4-5",
-                "fallbacks": ["gpt-5.1-codex-max", "gemini-3-pro"],
-                "description": "Complex analysis, architecture decisions",
-            },
-            "code-generation": {
-                "primary": "gpt-5.1-codex-max",
-                "fallbacks": ["claude-opus-4-5", "deepseek-v3.2"],
-                "description": "Code writing and refactoring",
-            },
-            "fast-response": {
-                "primary": "gemini-3-flash",
-                "fallbacks": ["qwen3-235b", "deepseek-v3.2"],
-                "description": "Quick tasks, simple queries",
-            },
-            "cost-optimized": {
-                "primary": "qwen3-235b",
-                "fallbacks": ["deepseek-v3.2"],
-                "description": "High-volume, non-critical tasks",
-            },
-            "self-hosted": {
-                "primary": "deepseek-v3.2",
-                "fallbacks": ["qwen3-235b"],
-                "description": "Privacy-sensitive, air-gapped",
-            },
-        }
-    }
+    return await _impl_get_usage(
+        project_id=project_id,
+        agent_id=agent_id,
+        period=period,
+    )
+
+
+@mcp.tool()
+async def count_tokens(
+    project_id: str,
+    agent_id: str,
+    text: str,
+    model: str | None = None,
+) -> dict[str, Any]:
+    """
+    Count tokens in text.
+
+    Args:
+        project_id: UUID of the project
+        agent_id: UUID of the agent
+        text: Text to count tokens in
+        model: Optional model for tokenizer selection
+
+    Returns:
+        Token count and estimation details
+    """
+    return await _impl_count_tokens(
+        project_id=project_id,
+        agent_id=agent_id,
+        text=text,
+        model=model,
+    )
+
+
+def main() -> None:
+    """Run the server."""
+    import uvicorn
+
+    settings = get_settings()
+    uvicorn.run(
+        "server:app",
+        host=settings.host,
+        port=settings.port,
+        reload=settings.debug,
+    )


 if __name__ == "__main__":
-    mcp.run()
+    main()
--- a/mcp-servers/llm-gateway/streaming.py
+++ b/mcp-servers/llm-gateway/streaming.py
@@ -0,0 +1,344 @@
+"""
+Streaming support for LLM Gateway.
+
+Provides async streaming wrappers for LiteLLM responses.
+"""
+
+import asyncio
+import json
+import logging
+import uuid
+from collections.abc import AsyncIterator
+from typing import Any
+
+from models import StreamChunk, UsageStats
+
+logger = logging.getLogger(__name__)
+
+
+class StreamAccumulator:
+    """
+    Accumulates streaming chunks for cost calculation.
+
+    Tracks:
+    - Full content for final response
+    - Token counts from chunks
+    - Timing information
+    """
+
+    def __init__(self, request_id: str | None = None) -> None:
+        """
+        Initialize accumulator.
+
+        Args:
+            request_id: Optional request ID for tracking
+        """
+        self.request_id = request_id or str(uuid.uuid4())
+        self.content_parts: list[str] = []
+        self.chunks_received = 0
+        self.prompt_tokens = 0
+        self.completion_tokens = 0
+        self.model: str | None = None
+        self.finish_reason: str | None = None
+        self._started_at: float | None = None
+        self._finished_at: float | None = None
+
+    @property
+    def content(self) -> str:
+        """Get accumulated content."""
+        return "".join(self.content_parts)
+
+    @property
+    def total_tokens(self) -> int:
+        """Get total token count."""
+        return self.prompt_tokens + self.completion_tokens
+
+    @property
+    def duration_seconds(self) -> float | None:
+        """Get stream duration in seconds."""
+        if self._started_at is None or self._finished_at is None:
+            return None
+        return self._finished_at - self._started_at
+
+    def start(self) -> None:
+        """Mark stream start."""
+        import time
+        self._started_at = time.time()
+
+    def finish(self) -> None:
+        """Mark stream finish."""
+        import time
+        self._finished_at = time.time()
+
+    def add_chunk(
+        self,
+        delta: str,
+        finish_reason: str | None = None,
+        model: str | None = None,
+        usage: dict[str, int] | None = None,
+    ) -> None:
+        """
+        Add a chunk to the accumulator.
+
+        Args:
+            delta: Content delta
+            finish_reason: Finish reason if this is the final chunk
+            model: Model name
+            usage: Usage stats if provided
+        """
+        if delta:
+            self.content_parts.append(delta)
+        self.chunks_received += 1
+
+        if finish_reason:
+            self.finish_reason = finish_reason
+
+        if model:
+            self.model = model
+
+        if usage:
+            self.prompt_tokens = usage.get("prompt_tokens", self.prompt_tokens)
+            self.completion_tokens = usage.get(
+                "completion_tokens", self.completion_tokens
+            )
+
+    def get_usage_stats(self, cost_usd: float = 0.0) -> UsageStats:
+        """Get usage statistics."""
+        return UsageStats(
+            prompt_tokens=self.prompt_tokens,
+            completion_tokens=self.completion_tokens,
+            total_tokens=self.total_tokens,
+            cost_usd=cost_usd,
+        )
+
+
+async def wrap_litellm_stream(
+    stream: AsyncIterator[Any],
+    accumulator: StreamAccumulator | None = None,
+) -> AsyncIterator[StreamChunk]:
+    """
+    Wrap a LiteLLM stream into StreamChunk objects.
+
+    Args:
+        stream: LiteLLM async stream
+        accumulator: Optional accumulator for tracking
+
+    Yields:
+        StreamChunk objects
+    """
+    if accumulator:
+        accumulator.start()
+
+    chunk_id = 0
+    try:
+        async for chunk in stream:
+            chunk_id += 1
+
+            # Extract data from LiteLLM chunk
+            delta = ""
+            finish_reason = None
+            usage = None
+            model = None
+
+            # Handle different chunk formats
+            if hasattr(chunk, "choices") and chunk.choices:
+                choice = chunk.choices[0]
+                if hasattr(choice, "delta"):
+                    delta = getattr(choice.delta, "content", "") or ""
+                finish_reason = getattr(choice, "finish_reason", None)
+
+            if hasattr(chunk, "model"):
+                model = chunk.model
+
+            if hasattr(chunk, "usage") and chunk.usage:
+                usage = {
+                    "prompt_tokens": getattr(chunk.usage, "prompt_tokens", 0),
+                    "completion_tokens": getattr(chunk.usage, "completion_tokens", 0),
+                }
+
+            # Update accumulator
+            if accumulator:
+                accumulator.add_chunk(
+                    delta=delta,
+                    finish_reason=finish_reason,
+                    model=model,
+                    usage=usage,
+                )
+
+            # Create StreamChunk
+            stream_chunk = StreamChunk(
+                id=f"{accumulator.request_id if accumulator else 'stream'}-{chunk_id}",
+                delta=delta,
+                finish_reason=finish_reason,
+            )
+
+            # Add usage on final chunk
+            if finish_reason and accumulator:
+                stream_chunk.usage = accumulator.get_usage_stats()
+
+            yield stream_chunk
+
+    finally:
+        if accumulator:
+            accumulator.finish()
+
+
+def format_sse_chunk(chunk: StreamChunk) -> str:
+    """
+    Format a StreamChunk as SSE data.
+
+    Args:
+        chunk: StreamChunk to format
+
+    Returns:
+        SSE-formatted string
+    """
+    data = {
+        "id": chunk.id,
+        "delta": chunk.delta,
+    }
+    if chunk.finish_reason:
+        data["finish_reason"] = chunk.finish_reason
+    if chunk.usage:
+        data["usage"] = chunk.usage.model_dump()
+
+    return f"data: {json.dumps(data)}\n\n"
+
+
+def format_sse_done() -> str:
+    """Format SSE done message."""
+    return "data: [DONE]\n\n"
+
+
+def format_sse_error(error: str, code: str | None = None) -> str:
+    """
+    Format an error as SSE data.
+
+    Args:
+        error: Error message
+        code: Error code
+
+    Returns:
+        SSE-formatted error string
+    """
+    data = {"error": error}
+    if code:
+        data["code"] = code
+    return f"data: {json.dumps(data)}\n\n"
+
+
+class StreamBuffer:
+    """
+    Buffer for streaming responses with backpressure handling.
+
+    Useful when producing chunks faster than they can be consumed.
+    """
+
+    def __init__(self, max_size: int = 100) -> None:
+        """
+        Initialize buffer.
+
+        Args:
+            max_size: Maximum buffer size
+        """
+        self._queue: asyncio.Queue[StreamChunk | None] = asyncio.Queue(maxsize=max_size)
+        self._done = False
+        self._error: Exception | None = None
+
+    async def put(self, chunk: StreamChunk) -> None:
+        """
+        Put a chunk in the buffer.
+
+        Args:
+            chunk: Chunk to buffer
+        """
+        if self._done:
+            raise RuntimeError("Buffer is closed")
+        await self._queue.put(chunk)
+
+    async def done(self) -> None:
+        """Signal that streaming is complete."""
+        self._done = True
+        await self._queue.put(None)
+
+    async def error(self, err: Exception) -> None:
+        """Signal an error occurred."""
+        self._error = err
+        self._done = True
+        await self._queue.put(None)
+
+    async def __aiter__(self) -> AsyncIterator[StreamChunk]:
+        """Iterate over buffered chunks."""
+        while True:
+            chunk = await self._queue.get()
+            if chunk is None:
+                if self._error:
+                    raise self._error
+                return
+            yield chunk
+
+
+async def stream_to_string(stream: AsyncIterator[StreamChunk]) -> tuple[str, UsageStats | None]:
+    """
+    Consume a stream and return full content.
+
+    Args:
+        stream: Stream to consume
+
+    Returns:
+        Tuple of (content, usage_stats)
+    """
+    parts: list[str] = []
+    usage: UsageStats | None = None
+
+    async for chunk in stream:
+        if chunk.delta:
+            parts.append(chunk.delta)
+        if chunk.usage:
+            usage = chunk.usage
+
+    return "".join(parts), usage
+
+
+async def merge_streams(
+    *streams: AsyncIterator[StreamChunk],
+) -> AsyncIterator[StreamChunk]:
+    """
+    Merge multiple streams into one.
+
+    Useful for parallel requests where results should be combined.
+
+    Args:
+        *streams: Streams to merge
+
+    Yields:
+        Chunks from all streams in arrival order
+    """
+    pending: set[asyncio.Task[tuple[int, StreamChunk | None]]] = set()
+
+    async def next_chunk(
+        idx: int, stream: AsyncIterator[StreamChunk]
+    ) -> tuple[int, StreamChunk | None]:
+        try:
+            return idx, await stream.__anext__()
+        except StopAsyncIteration:
+            return idx, None
+
+    # Start initial tasks
+    active_streams = list(streams)
+    for idx, stream in enumerate(active_streams):
+        task = asyncio.create_task(next_chunk(idx, stream))
+        pending.add(task)
+
+    while pending:
+        done, pending = await asyncio.wait(
+            pending, return_when=asyncio.FIRST_COMPLETED
+        )
+
+        for task in done:
+            idx, chunk = task.result()
+            if chunk is not None:
+                yield chunk
+                # Schedule next chunk from this stream
+                new_task = asyncio.create_task(next_chunk(idx, active_streams[idx]))
+                pending.add(new_task)
--- a/mcp-servers/llm-gateway/tests/init.py
+++ b/mcp-servers/llm-gateway/tests/init.py
@@ -0,0 +1 @@
+"""Tests for LLM Gateway MCP Server."""
--- a/mcp-servers/llm-gateway/tests/conftest.py
+++ b/mcp-servers/llm-gateway/tests/conftest.py
@@ -0,0 +1,204 @@
+"""
+Pytest fixtures for LLM Gateway tests.
+"""
+
+import os
+import sys
+from collections.abc import AsyncIterator, Iterator
+from typing import Any
+from unittest.mock import MagicMock, patch
+
+import fakeredis.aioredis
+import pytest
+
+# Add parent directory to path for imports
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from config import Settings
+from cost_tracking import CostTracker, reset_cost_tracker
+from failover import CircuitBreakerRegistry, reset_circuit_registry
+from providers import LLMProvider, reset_provider
+from routing import ModelRouter, reset_model_router
+
+
+@pytest.fixture
+def test_settings() -> Settings:
+    """Create test settings with mock API keys."""
+    return Settings(
+        host="127.0.0.1",
+        port=8001,
+        debug=True,
+        redis_url="redis://localhost:6379/0",
+        anthropic_api_key="test-anthropic-key",
+        openai_api_key="test-openai-key",
+        google_api_key="test-google-key",
+        litellm_timeout=30,
+        litellm_max_retries=2,
+        litellm_cache_enabled=False,
+        cost_tracking_enabled=True,
+        circuit_failure_threshold=3,
+        circuit_recovery_timeout=10,
+    )
+
+
+@pytest.fixture
+def settings_no_providers() -> Settings:
+    """Create settings with no providers configured."""
+    return Settings(
+        host="127.0.0.1",
+        port=8001,
+        debug=False,
+        redis_url="redis://localhost:6379/0",
+        anthropic_api_key=None,
+        openai_api_key=None,
+        google_api_key=None,
+        alibaba_api_key=None,
+        deepseek_api_key=None,
+    )
+
+
+@pytest.fixture
+def fake_redis() -> fakeredis.aioredis.FakeRedis:
+    """Create a fake Redis instance for testing."""
+    return fakeredis.aioredis.FakeRedis(decode_responses=True)
+
+
+@pytest.fixture
+async def cost_tracker(
+    fake_redis: fakeredis.aioredis.FakeRedis,
+    test_settings: Settings,
+) -> AsyncIterator[CostTracker]:
+    """Create a cost tracker with fake Redis."""
+    reset_cost_tracker()
+    tracker = CostTracker(redis_client=fake_redis, settings=test_settings)
+    yield tracker
+    await tracker.close()
+    reset_cost_tracker()
+
+
+@pytest.fixture
+def circuit_registry(test_settings: Settings) -> Iterator[CircuitBreakerRegistry]:
+    """Create a circuit breaker registry for testing."""
+    reset_circuit_registry()
+    registry = CircuitBreakerRegistry(settings=test_settings)
+    yield registry
+    reset_circuit_registry()
+
+
+@pytest.fixture
+def model_router(
+    test_settings: Settings,
+    circuit_registry: CircuitBreakerRegistry,
+) -> Iterator[ModelRouter]:
+    """Create a model router for testing."""
+    reset_model_router()
+    router = ModelRouter(settings=test_settings, circuit_registry=circuit_registry)
+    yield router
+    reset_model_router()
+
+
+@pytest.fixture
+def llm_provider(test_settings: Settings) -> Iterator[LLMProvider]:
+    """Create an LLM provider for testing."""
+    reset_provider()
+    provider = LLMProvider(settings=test_settings)
+    yield provider
+    reset_provider()
+
+
+@pytest.fixture
+def mock_litellm_response() -> dict[str, Any]:
+    """Create a mock LiteLLM response."""
+    return {
+        "id": "test-response-id",
+        "model": "claude-opus-4",
+        "choices": [
+            {
+                "index": 0,
+                "message": {
+                    "role": "assistant",
+                    "content": "This is a test response.",
+                },
+                "finish_reason": "stop",
+            }
+        ],
+        "usage": {
+            "prompt_tokens": 10,
+            "completion_tokens": 20,
+            "total_tokens": 30,
+        },
+    }
+
+
+@pytest.fixture
+def mock_completion_response() -> MagicMock:
+    """Create a mock completion response object."""
+    response = MagicMock()
+    response.id = "test-response-id"
+    response.model = "claude-opus-4"
+
+    choice = MagicMock()
+    choice.index = 0
+    choice.message = MagicMock()
+    choice.message.content = "This is a test response."
+    choice.finish_reason = "stop"
+    response.choices = [choice]
+
+    response.usage = MagicMock()
+    response.usage.prompt_tokens = 10
+    response.usage.completion_tokens = 20
+    response.usage.total_tokens = 30
+
+    return response
+
+
+@pytest.fixture
+def sample_messages() -> list[dict[str, str]]:
+    """Sample chat messages for testing."""
+    return [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "Hello, how are you?"},
+    ]
+
+
+@pytest.fixture
+def sample_project_id() -> str:
+    """Sample project ID."""
+    return "proj-12345678-1234-1234-1234-123456789abc"
+
+
+@pytest.fixture
+def sample_agent_id() -> str:
+    """Sample agent ID."""
+    return "agent-87654321-4321-4321-4321-cba987654321"
+
+
+@pytest.fixture
+def sample_session_id() -> str:
+    """Sample session ID."""
+    return "session-11111111-2222-3333-4444-555555555555"
+
+
+# Reset all global state after each test
+@pytest.fixture(autouse=True)
+def reset_globals() -> Iterator[None]:
+    """Reset all global state after each test."""
+    yield
+    reset_cost_tracker()
+    reset_circuit_registry()
+    reset_model_router()
+    reset_provider()
+
+
+# Mock environment variables for tests
+# Note: Not autouse=True to avoid affecting default value tests
+@pytest.fixture
+def mock_env_vars() -> Iterator[None]:
+    """Set test environment variables."""
+    env_vars = {
+        "LLM_GATEWAY_HOST": "127.0.0.1",
+        "LLM_GATEWAY_PORT": "8001",
+        "LLM_GATEWAY_DEBUG": "true",
+    }
+    with patch.dict(os.environ, env_vars, clear=False):
+        yield
--- a/mcp-servers/llm-gateway/tests/test_config.py
+++ b/mcp-servers/llm-gateway/tests/test_config.py
@@ -0,0 +1,200 @@
+"""
+Tests for config module.
+"""
+
+import os
+from unittest.mock import patch
+
+import pytest
+
+from config import Settings, get_settings
+
+
+class TestSettings:
+    """Tests for Settings class."""
+
+    def test_default_values(self) -> None:
+        """Test default configuration values."""
+        settings = Settings()
+
+        assert settings.host == "0.0.0.0"
+        assert settings.port == 8001
+        assert settings.debug is False
+        assert settings.redis_url == "redis://localhost:6379/0"
+        assert settings.litellm_timeout == 120
+        assert settings.circuit_failure_threshold == 5
+
+    def test_custom_values(self) -> None:
+        """Test custom configuration values."""
+        settings = Settings(
+            host="127.0.0.1",
+            port=9000,
+            debug=True,
+            redis_url="redis://custom:6380/1",
+            litellm_timeout=60,
+        )
+
+        assert settings.host == "127.0.0.1"
+        assert settings.port == 9000
+        assert settings.debug is True
+        assert settings.redis_url == "redis://custom:6380/1"
+        assert settings.litellm_timeout == 60
+
+    def test_port_validation_valid(self) -> None:
+        """Test valid port numbers."""
+        settings = Settings(port=1)
+        assert settings.port == 1
+
+        settings = Settings(port=65535)
+        assert settings.port == 65535
+
+        settings = Settings(port=8080)
+        assert settings.port == 8080
+
+    def test_port_validation_invalid(self) -> None:
+        """Test invalid port numbers."""
+        with pytest.raises(ValueError, match="Port must be between"):
+            Settings(port=0)
+
+        with pytest.raises(ValueError, match="Port must be between"):
+            Settings(port=65536)
+
+        with pytest.raises(ValueError, match="Port must be between"):
+            Settings(port=-1)
+
+    def test_ttl_validation_valid(self) -> None:
+        """Test valid TTL values."""
+        settings = Settings(redis_ttl_hours=1)
+        assert settings.redis_ttl_hours == 1
+
+        settings = Settings(redis_ttl_hours=168)  # 1 week
+        assert settings.redis_ttl_hours == 168
+
+    def test_ttl_validation_invalid(self) -> None:
+        """Test invalid TTL values."""
+        with pytest.raises(ValueError, match="Redis TTL must be positive"):
+            Settings(redis_ttl_hours=0)
+
+        with pytest.raises(ValueError, match="Redis TTL must be positive"):
+            Settings(redis_ttl_hours=-1)
+
+    def test_failure_threshold_validation(self) -> None:
+        """Test circuit failure threshold validation."""
+        settings = Settings(circuit_failure_threshold=1)
+        assert settings.circuit_failure_threshold == 1
+
+        settings = Settings(circuit_failure_threshold=100)
+        assert settings.circuit_failure_threshold == 100
+
+        with pytest.raises(ValueError, match="Failure threshold must be between"):
+            Settings(circuit_failure_threshold=0)
+
+        with pytest.raises(ValueError, match="Failure threshold must be between"):
+            Settings(circuit_failure_threshold=101)
+
+    def test_timeout_validation(self) -> None:
+        """Test timeout validation."""
+        settings = Settings(litellm_timeout=1)
+        assert settings.litellm_timeout == 1
+
+        settings = Settings(litellm_timeout=600)
+        assert settings.litellm_timeout == 600
+
+        with pytest.raises(ValueError, match="Timeout must be between"):
+            Settings(litellm_timeout=0)
+
+        with pytest.raises(ValueError, match="Timeout must be between"):
+            Settings(litellm_timeout=601)
+
+    def test_get_available_providers_none(self) -> None:
+        """Test getting available providers with none configured."""
+        settings = Settings()
+        providers = settings.get_available_providers()
+        assert providers == []
+
+    def test_get_available_providers_some(self) -> None:
+        """Test getting available providers with some configured."""
+        settings = Settings(
+            anthropic_api_key="test-key",
+            openai_api_key="test-key",
+        )
+        providers = settings.get_available_providers()
+
+        assert "anthropic" in providers
+        assert "openai" in providers
+        assert "google" not in providers
+        assert len(providers) == 2
+
+    def test_get_available_providers_all(self) -> None:
+        """Test getting available providers with all configured."""
+        settings = Settings(
+            anthropic_api_key="test-key",
+            openai_api_key="test-key",
+            google_api_key="test-key",
+            alibaba_api_key="test-key",
+            deepseek_api_key="test-key",
+        )
+        providers = settings.get_available_providers()
+
+        assert len(providers) == 5
+        assert "anthropic" in providers
+        assert "openai" in providers
+        assert "google" in providers
+        assert "alibaba" in providers
+        assert "deepseek" in providers
+
+    def test_has_any_provider_false(self) -> None:
+        """Test has_any_provider when none configured."""
+        settings = Settings()
+        assert settings.has_any_provider() is False
+
+    def test_has_any_provider_true(self) -> None:
+        """Test has_any_provider when at least one configured."""
+        settings = Settings(anthropic_api_key="test-key")
+        assert settings.has_any_provider() is True
+
+    def test_deepseek_base_url_counts_as_provider(self) -> None:
+        """Test that DeepSeek base URL alone counts as provider."""
+        settings = Settings(deepseek_base_url="http://localhost:8000")
+        providers = settings.get_available_providers()
+        assert "deepseek" in providers
+
+
+class TestGetSettings:
+    """Tests for get_settings function."""
+
+    def test_get_settings_returns_settings(self) -> None:
+        """Test that get_settings returns a Settings instance."""
+        # Clear the cache first
+        get_settings.cache_clear()
+
+        settings = get_settings()
+        assert isinstance(settings, Settings)
+
+    def test_get_settings_is_cached(self) -> None:
+        """Test that get_settings returns cached instance."""
+        get_settings.cache_clear()
+
+        settings1 = get_settings()
+        settings2 = get_settings()
+
+        assert settings1 is settings2
+
+    def test_env_var_override(self) -> None:
+        """Test that environment variables override defaults."""
+        get_settings.cache_clear()
+
+        with patch.dict(
+            os.environ,
+            {
+                "LLM_GATEWAY_HOST": "192.168.1.1",
+                "LLM_GATEWAY_PORT": "9999",
+                "LLM_GATEWAY_DEBUG": "true",
+            },
+        ):
+            get_settings.cache_clear()
+            settings = get_settings()
+
+            assert settings.host == "192.168.1.1"
+            assert settings.port == 9999
+            assert settings.debug is True
--- a/mcp-servers/llm-gateway/tests/test_cost_tracking.py
+++ b/mcp-servers/llm-gateway/tests/test_cost_tracking.py
@@ -0,0 +1,417 @@
+"""
+Tests for cost_tracking module.
+"""
+
+import asyncio
+
+import fakeredis.aioredis
+import pytest
+
+from config import Settings
+from cost_tracking import (
+    CostTracker,
+    calculate_cost,
+    close_cost_tracker,
+    get_cost_tracker,
+    reset_cost_tracker,
+)
+
+
+@pytest.fixture
+def tracker_settings() -> Settings:
+    """Settings for cost tracker tests."""
+    return Settings(
+        redis_url="redis://localhost:6379/0",
+        redis_prefix="test_llm_gateway",
+        cost_tracking_enabled=True,
+        cost_alert_threshold=100.0,
+        default_budget_limit=1000.0,
+    )
+
+
+@pytest.fixture
+def fake_redis() -> fakeredis.aioredis.FakeRedis:
+    """Create fake Redis for testing."""
+    return fakeredis.aioredis.FakeRedis(decode_responses=True)
+
+
+@pytest.fixture
+def tracker(
+    fake_redis: fakeredis.aioredis.FakeRedis,
+    tracker_settings: Settings,
+) -> CostTracker:
+    """Create cost tracker with fake Redis."""
+    return CostTracker(redis_client=fake_redis, settings=tracker_settings)
+
+
+class TestCalculateCost:
+    """Tests for calculate_cost function."""
+
+    def test_calculate_cost_known_model(self) -> None:
+        """Test calculating cost for known model."""
+        # claude-opus-4: $15/1M input, $75/1M output
+        cost = calculate_cost(
+            model="claude-opus-4",
+            prompt_tokens=1000,
+            completion_tokens=500,
+        )
+
+        # 1000/1M * 15 + 500/1M * 75 = 0.015 + 0.0375 = 0.0525
+        assert cost == pytest.approx(0.0525, rel=0.001)
+
+    def test_calculate_cost_unknown_model(self) -> None:
+        """Test calculating cost for unknown model."""
+        cost = calculate_cost(
+            model="unknown-model",
+            prompt_tokens=1000,
+            completion_tokens=500,
+        )
+
+        assert cost == 0.0
+
+    def test_calculate_cost_zero_tokens(self) -> None:
+        """Test calculating cost with zero tokens."""
+        cost = calculate_cost(
+            model="claude-opus-4",
+            prompt_tokens=0,
+            completion_tokens=0,
+        )
+
+        assert cost == 0.0
+
+    def test_calculate_cost_large_token_counts(self) -> None:
+        """Test calculating cost with large token counts."""
+        cost = calculate_cost(
+            model="claude-opus-4",
+            prompt_tokens=1_000_000,
+            completion_tokens=500_000,
+        )
+
+        # 1M * 15/1M + 500K * 75/1M = 15 + 37.5 = 52.5
+        assert cost == pytest.approx(52.5, rel=0.001)
+
+
+class TestCostTracker:
+    """Tests for CostTracker class."""
+
+    def test_record_usage(self, tracker: CostTracker) -> None:
+        """Test recording usage."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+            )
+        )
+
+        # Verify by getting usage report
+        report = asyncio.run(
+            tracker.get_project_usage("proj-123", period="day")
+        )
+
+        assert report.total_requests == 1
+        assert report.total_cost_usd == pytest.approx(0.01, rel=0.01)
+
+    def test_record_usage_disabled(self, tracker_settings: Settings) -> None:
+        """Test recording is skipped when disabled."""
+        settings = Settings(**{
+            **tracker_settings.model_dump(),
+            "cost_tracking_enabled": False,
+        })
+        fake_redis = fakeredis.aioredis.FakeRedis(decode_responses=True)
+        disabled_tracker = CostTracker(redis_client=fake_redis, settings=settings)
+
+        # This should not raise and should not record
+        asyncio.run(
+            disabled_tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+            )
+        )
+
+        # Usage should be empty
+        report = asyncio.run(
+            disabled_tracker.get_project_usage("proj-123", period="day")
+        )
+        assert report.total_requests == 0
+
+    def test_record_usage_with_session(self, tracker: CostTracker) -> None:
+        """Test recording usage with session ID."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+                session_id="session-789",
+            )
+        )
+
+        # Verify session usage
+        session_usage = asyncio.run(
+            tracker.get_session_usage("session-789")
+        )
+
+        assert session_usage["session_id"] == "session-789"
+        assert session_usage["total_cost_usd"] == pytest.approx(0.01, rel=0.01)
+
+    def test_get_project_usage_empty(self, tracker: CostTracker) -> None:
+        """Test getting usage for project with no data."""
+        report = asyncio.run(
+            tracker.get_project_usage("nonexistent-project", period="day")
+        )
+
+        assert report.entity_id == "nonexistent-project"
+        assert report.entity_type == "project"
+        assert report.total_requests == 0
+        assert report.total_cost_usd == 0.0
+
+    def test_get_project_usage_multiple_models(self, tracker: CostTracker) -> None:
+        """Test usage tracking across multiple models."""
+        # Record usage for different models
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+            )
+        )
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="gpt-4.1",
+                prompt_tokens=200,
+                completion_tokens=100,
+                cost_usd=0.02,
+            )
+        )
+
+        report = asyncio.run(
+            tracker.get_project_usage("proj-123", period="day")
+        )
+
+        assert report.total_requests == 2
+        assert len(report.by_model) == 2
+        assert "claude-opus-4" in report.by_model
+        assert "gpt-4.1" in report.by_model
+
+    def test_get_agent_usage(self, tracker: CostTracker) -> None:
+        """Test getting agent usage."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+            )
+        )
+
+        report = asyncio.run(
+            tracker.get_agent_usage("agent-456", period="day")
+        )
+
+        assert report.entity_id == "agent-456"
+        assert report.entity_type == "agent"
+        assert report.total_requests == 1
+
+    def test_usage_periods(self, tracker: CostTracker) -> None:
+        """Test different usage periods."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=100,
+                completion_tokens=50,
+                cost_usd=0.01,
+            )
+        )
+
+        # Check different periods
+        hour_report = asyncio.run(
+            tracker.get_project_usage("proj-123", period="hour")
+        )
+        day_report = asyncio.run(
+            tracker.get_project_usage("proj-123", period="day")
+        )
+        month_report = asyncio.run(
+            tracker.get_project_usage("proj-123", period="month")
+        )
+
+        assert hour_report.period == "hour"
+        assert day_report.period == "day"
+        assert month_report.period == "month"
+
+        # All should have the same data
+        assert hour_report.total_requests == 1
+        assert day_report.total_requests == 1
+        assert month_report.total_requests == 1
+
+    def test_check_budget_within(self, tracker: CostTracker) -> None:
+        """Test budget check when within limit."""
+        # Record some usage
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=1000,
+                completion_tokens=500,
+                cost_usd=50.0,
+            )
+        )
+
+        within, current, limit = asyncio.run(
+            tracker.check_budget("proj-123", budget_limit=100.0)
+        )
+
+        assert within is True
+        assert current == pytest.approx(50.0, rel=0.01)
+        assert limit == 100.0
+
+    def test_check_budget_exceeded(self, tracker: CostTracker) -> None:
+        """Test budget check when exceeded."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=1000,
+                completion_tokens=500,
+                cost_usd=150.0,
+            )
+        )
+
+        within, current, limit = asyncio.run(
+            tracker.check_budget("proj-123", budget_limit=100.0)
+        )
+
+        assert within is False
+        assert current >= limit
+
+    def test_check_budget_default_limit(self, tracker: CostTracker) -> None:
+        """Test budget check with default limit."""
+        within, current, limit = asyncio.run(
+            tracker.check_budget("proj-123")
+        )
+
+        assert limit == 1000.0  # Default from settings
+
+    def test_estimate_request_cost_known_model(self, tracker: CostTracker) -> None:
+        """Test estimating cost for known model."""
+        cost = asyncio.run(
+            tracker.estimate_request_cost(
+                model="claude-opus-4",
+                prompt_tokens=1000,
+                max_completion_tokens=500,
+            )
+        )
+
+        # 1000/1M * 15 + 500/1M * 75 = 0.015 + 0.0375 = 0.0525
+        assert cost == pytest.approx(0.0525, rel=0.01)
+
+    def test_estimate_request_cost_unknown_model(self, tracker: CostTracker) -> None:
+        """Test estimating cost for unknown model."""
+        cost = asyncio.run(
+            tracker.estimate_request_cost(
+                model="unknown-model",
+                prompt_tokens=1000,
+                max_completion_tokens=500,
+            )
+        )
+
+        # Uses fallback estimate
+        assert cost > 0
+
+    def test_should_alert_below_threshold(self, tracker: CostTracker) -> None:
+        """Test alert check when below threshold."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=1000,
+                completion_tokens=500,
+                cost_usd=50.0,
+            )
+        )
+
+        should_alert, current = asyncio.run(
+            tracker.should_alert("proj-123", threshold=100.0)
+        )
+
+        assert should_alert is False
+        assert current == pytest.approx(50.0, rel=0.01)
+
+    def test_should_alert_above_threshold(self, tracker: CostTracker) -> None:
+        """Test alert check when above threshold."""
+        asyncio.run(
+            tracker.record_usage(
+                project_id="proj-123",
+                agent_id="agent-456",
+                model="claude-opus-4",
+                prompt_tokens=1000,
+                completion_tokens=500,
+                cost_usd=150.0,
+            )
+        )
+
+        should_alert, current = asyncio.run(
+            tracker.should_alert("proj-123", threshold=100.0)
+        )
+
+        assert should_alert is True
+
+    def test_close(self, tracker: CostTracker) -> None:
+        """Test closing tracker."""
+        asyncio.run(tracker.close())
+        assert tracker._redis is None
+
+
+class TestGlobalTracker:
+    """Tests for global tracker functions."""
+
+    def test_get_cost_tracker(self) -> None:
+        """Test getting global tracker."""
+        reset_cost_tracker()
+        tracker = get_cost_tracker()
+        assert isinstance(tracker, CostTracker)
+
+    def test_get_cost_tracker_singleton(self) -> None:
+        """Test tracker is singleton."""
+        reset_cost_tracker()
+        tracker1 = get_cost_tracker()
+        tracker2 = get_cost_tracker()
+        assert tracker1 is tracker2
+
+    def test_reset_cost_tracker(self) -> None:
+        """Test resetting global tracker."""
+        reset_cost_tracker()
+        tracker1 = get_cost_tracker()
+        reset_cost_tracker()
+        tracker2 = get_cost_tracker()
+        assert tracker1 is not tracker2
+
+    def test_close_cost_tracker(self) -> None:
+        """Test closing global tracker."""
+        reset_cost_tracker()
+        _ = get_cost_tracker()
+        asyncio.run(close_cost_tracker())
+        # Getting again should create a new one
+        tracker2 = get_cost_tracker()
+        assert tracker2 is not None
--- a/mcp-servers/llm-gateway/tests/test_exceptions.py
+++ b/mcp-servers/llm-gateway/tests/test_exceptions.py
@@ -0,0 +1,377 @@
+"""
+Tests for exceptions module.
+"""
+
+
+from exceptions import (
+    AllProvidersFailedError,
+    CircuitOpenError,
+    ConfigurationError,
+    ContextTooLongError,
+    CostLimitExceededError,
+    ErrorCode,
+    InvalidModelError,
+    InvalidModelGroupError,
+    LLMGatewayError,
+    ModelNotAvailableError,
+    ProviderError,
+    RateLimitError,
+    StreamError,
+    TokenLimitExceededError,
+)
+
+
+class TestErrorCode:
+    """Tests for ErrorCode enum."""
+
+    def test_error_code_values(self) -> None:
+        """Test error code values."""
+        assert ErrorCode.UNKNOWN_ERROR.value == "LLM_UNKNOWN_ERROR"
+        assert ErrorCode.PROVIDER_ERROR.value == "LLM_PROVIDER_ERROR"
+        assert ErrorCode.CIRCUIT_OPEN.value == "LLM_CIRCUIT_OPEN"
+        assert ErrorCode.COST_LIMIT_EXCEEDED.value == "LLM_COST_LIMIT_EXCEEDED"
+
+
+class TestLLMGatewayError:
+    """Tests for LLMGatewayError base class."""
+
+    def test_basic_error(self) -> None:
+        """Test basic error creation."""
+        error = LLMGatewayError("Something went wrong")
+        assert str(error) == "[LLM_UNKNOWN_ERROR] Something went wrong"
+        assert error.message == "Something went wrong"
+        assert error.code == ErrorCode.UNKNOWN_ERROR
+        assert error.details == {}
+        assert error.cause is None
+
+    def test_error_with_code(self) -> None:
+        """Test error with custom code."""
+        error = LLMGatewayError(
+            "Provider failed",
+            code=ErrorCode.PROVIDER_ERROR,
+        )
+        assert error.code == ErrorCode.PROVIDER_ERROR
+
+    def test_error_with_details(self) -> None:
+        """Test error with details."""
+        error = LLMGatewayError(
+            "Error",
+            details={"key": "value"},
+        )
+        assert error.details == {"key": "value"}
+
+    def test_error_with_cause(self) -> None:
+        """Test error with cause exception."""
+        cause = ValueError("Original error")
+        error = LLMGatewayError("Wrapped error", cause=cause)
+        assert error.cause is cause
+
+    def test_to_dict(self) -> None:
+        """Test converting error to dict."""
+        error = LLMGatewayError(
+            "Test error",
+            code=ErrorCode.INVALID_REQUEST,
+            details={"field": "value"},
+        )
+        result = error.to_dict()
+
+        assert result["error"] == "LLM_INVALID_REQUEST"
+        assert result["message"] == "Test error"
+        assert result["details"] == {"field": "value"}
+
+    def test_to_dict_no_details(self) -> None:
+        """Test to_dict without details."""
+        error = LLMGatewayError("Test error")
+        result = error.to_dict()
+
+        assert "details" not in result
+
+    def test_repr(self) -> None:
+        """Test error repr."""
+        error = LLMGatewayError("Test", details={"key": "val"})
+        repr_str = repr(error)
+
+        assert "LLMGatewayError" in repr_str
+        assert "Test" in repr_str
+
+
+class TestProviderError:
+    """Tests for ProviderError."""
+
+    def test_basic_provider_error(self) -> None:
+        """Test basic provider error."""
+        error = ProviderError(
+            message="API call failed",
+            provider="anthropic",
+        )
+
+        assert error.provider == "anthropic"
+        assert error.model is None
+        assert error.status_code is None
+        assert error.code == ErrorCode.PROVIDER_ERROR
+        assert "provider" in error.details
+
+    def test_provider_error_with_model(self) -> None:
+        """Test provider error with model info."""
+        error = ProviderError(
+            message="Model not found",
+            provider="openai",
+            model="gpt-5",
+            status_code=404,
+        )
+
+        assert error.model == "gpt-5"
+        assert error.status_code == 404
+        assert error.details["model"] == "gpt-5"
+        assert error.details["status_code"] == 404
+
+    def test_provider_error_with_cause(self) -> None:
+        """Test provider error with cause."""
+        cause = ConnectionError("Network down")
+        error = ProviderError(
+            message="Connection failed",
+            provider="google",
+            cause=cause,
+        )
+
+        assert error.cause is cause
+
+
+class TestRateLimitError:
+    """Tests for RateLimitError."""
+
+    def test_internal_rate_limit(self) -> None:
+        """Test internal rate limit error."""
+        error = RateLimitError(
+            message="Too many requests",
+            retry_after=60,
+        )
+
+        assert error.code == ErrorCode.RATE_LIMIT_EXCEEDED
+        assert error.provider is None
+        assert error.retry_after == 60
+        assert error.details["retry_after_seconds"] == 60
+
+    def test_provider_rate_limit(self) -> None:
+        """Test provider rate limit error."""
+        error = RateLimitError(
+            message="OpenAI rate limit",
+            provider="openai",
+            retry_after=30,
+        )
+
+        assert error.code == ErrorCode.PROVIDER_RATE_LIMIT
+        assert error.provider == "openai"
+        assert error.details["provider"] == "openai"
+
+
+class TestCircuitOpenError:
+    """Tests for CircuitOpenError."""
+
+    def test_circuit_open_error(self) -> None:
+        """Test circuit open error."""
+        error = CircuitOpenError(
+            provider="anthropic",
+            recovery_time=45,
+        )
+
+        assert error.provider == "anthropic"
+        assert error.recovery_time == 45
+        assert error.code == ErrorCode.CIRCUIT_OPEN
+        assert "Circuit breaker open" in error.message
+        assert error.details["recovery_time_seconds"] == 45
+
+    def test_circuit_open_no_recovery_time(self) -> None:
+        """Test circuit open without recovery time."""
+        error = CircuitOpenError(provider="openai")
+
+        assert error.recovery_time is None
+        assert "recovery_time_seconds" not in error.details
+
+
+class TestCostLimitExceededError:
+    """Tests for CostLimitExceededError."""
+
+    def test_project_cost_limit(self) -> None:
+        """Test project cost limit error."""
+        error = CostLimitExceededError(
+            entity_type="project",
+            entity_id="proj-123",
+            current_cost=150.0,
+            limit=100.0,
+        )
+
+        assert error.entity_type == "project"
+        assert error.entity_id == "proj-123"
+        assert error.current_cost == 150.0
+        assert error.limit == 100.0
+        assert error.code == ErrorCode.COST_LIMIT_EXCEEDED
+        assert "$150.00" in error.message
+        assert "$100.00" in error.message
+
+    def test_agent_cost_limit(self) -> None:
+        """Test agent cost limit error."""
+        error = CostLimitExceededError(
+            entity_type="agent",
+            entity_id="agent-456",
+            current_cost=50.0,
+            limit=25.0,
+        )
+
+        assert error.entity_type == "agent"
+        assert error.details["entity_type"] == "agent"
+
+
+class TestInvalidModelGroupError:
+    """Tests for InvalidModelGroupError."""
+
+    def test_invalid_group_error(self) -> None:
+        """Test invalid model group error."""
+        error = InvalidModelGroupError(
+            model_group="invalid_group",
+            available_groups=["reasoning", "code", "fast"],
+        )
+
+        assert error.model_group == "invalid_group"
+        assert error.available_groups == ["reasoning", "code", "fast"]
+        assert error.code == ErrorCode.INVALID_MODEL_GROUP
+        assert "invalid_group" in error.message
+
+    def test_invalid_group_no_available(self) -> None:
+        """Test invalid group without available list."""
+        error = InvalidModelGroupError(model_group="unknown")
+
+        assert error.available_groups is None
+        assert "available_groups" not in error.details
+
+
+class TestInvalidModelError:
+    """Tests for InvalidModelError."""
+
+    def test_invalid_model_error(self) -> None:
+        """Test invalid model error."""
+        error = InvalidModelError(
+            model="gpt-99",
+            reason="Model does not exist",
+        )
+
+        assert error.model == "gpt-99"
+        assert error.code == ErrorCode.INVALID_MODEL
+        assert "gpt-99" in error.message
+        assert "Model does not exist" in error.message
+
+    def test_invalid_model_no_reason(self) -> None:
+        """Test invalid model without reason."""
+        error = InvalidModelError(model="unknown-model")
+
+        assert "reason" not in error.details
+
+
+class TestModelNotAvailableError:
+    """Tests for ModelNotAvailableError."""
+
+    def test_model_not_available(self) -> None:
+        """Test model not available error."""
+        error = ModelNotAvailableError(
+            model="claude-opus-4",
+            provider="anthropic",
+        )
+
+        assert error.model == "claude-opus-4"
+        assert error.provider == "anthropic"
+        assert error.code == ErrorCode.MODEL_NOT_AVAILABLE
+        assert "not configured" in error.message
+
+
+class TestAllProvidersFailedError:
+    """Tests for AllProvidersFailedError."""
+
+    def test_all_providers_failed(self) -> None:
+        """Test all providers failed error."""
+        errors = [
+            {"model": "claude-opus-4", "error": "Rate limited"},
+            {"model": "gpt-4.1", "error": "Timeout"},
+        ]
+        error = AllProvidersFailedError(
+            model_group="reasoning",
+            attempted_models=["claude-opus-4", "gpt-4.1"],
+            errors=errors,
+        )
+
+        assert error.model_group == "reasoning"
+        assert error.attempted_models == ["claude-opus-4", "gpt-4.1"]
+        assert error.errors == errors
+        assert error.code == ErrorCode.ALL_PROVIDERS_FAILED
+
+
+class TestStreamError:
+    """Tests for StreamError."""
+
+    def test_stream_error(self) -> None:
+        """Test stream error."""
+        cause = OSError("Connection reset")
+        error = StreamError(
+            message="Stream interrupted",
+            chunks_received=10,
+            cause=cause,
+        )
+
+        assert error.chunks_received == 10
+        assert error.cause is cause
+        assert error.code == ErrorCode.STREAM_ERROR
+
+
+class TestTokenLimitExceededError:
+    """Tests for TokenLimitExceededError."""
+
+    def test_token_limit_exceeded(self) -> None:
+        """Test token limit exceeded error."""
+        error = TokenLimitExceededError(
+            model="claude-haiku",
+            token_count=10000,
+            limit=8192,
+        )
+
+        assert error.model == "claude-haiku"
+        assert error.token_count == 10000
+        assert error.limit == 8192
+        assert error.code == ErrorCode.TOKEN_LIMIT_EXCEEDED
+
+
+class TestContextTooLongError:
+    """Tests for ContextTooLongError."""
+
+    def test_context_too_long(self) -> None:
+        """Test context too long error."""
+        error = ContextTooLongError(
+            model="gpt-4.1-mini",
+            context_length=150000,
+            max_context=100000,
+        )
+
+        assert error.model == "gpt-4.1-mini"
+        assert error.context_length == 150000
+        assert error.max_context == 100000
+        assert error.code == ErrorCode.CONTEXT_TOO_LONG
+
+
+class TestConfigurationError:
+    """Tests for ConfigurationError."""
+
+    def test_configuration_error(self) -> None:
+        """Test configuration error."""
+        error = ConfigurationError(
+            message="Missing API key",
+            config_key="ANTHROPIC_API_KEY",
+        )
+
+        assert error.config_key == "ANTHROPIC_API_KEY"
+        assert error.code == ErrorCode.CONFIGURATION_ERROR
+        assert error.details["config_key"] == "ANTHROPIC_API_KEY"
+
+    def test_configuration_error_no_key(self) -> None:
+        """Test configuration error without key."""
+        error = ConfigurationError(message="Invalid configuration")
+
+        assert error.config_key is None
+        assert "config_key" not in error.details
--- a/mcp-servers/llm-gateway/tests/test_failover.py
+++ b/mcp-servers/llm-gateway/tests/test_failover.py
@@ -0,0 +1,407 @@
+"""
+Tests for failover module (circuit breaker).
+"""
+
+import asyncio
+import time
+
+import pytest
+
+from config import Settings
+from exceptions import CircuitOpenError
+from failover import (
+    CircuitBreaker,
+    CircuitBreakerRegistry,
+    CircuitState,
+    CircuitStats,
+    get_circuit_registry,
+    reset_circuit_registry,
+)
+
+
+class TestCircuitState:
+    """Tests for CircuitState enum."""
+
+    def test_circuit_states(self) -> None:
+        """Test circuit state values."""
+        assert CircuitState.CLOSED.value == "closed"
+        assert CircuitState.OPEN.value == "open"
+        assert CircuitState.HALF_OPEN.value == "half_open"
+
+
+class TestCircuitStats:
+    """Tests for CircuitStats dataclass."""
+
+    def test_default_stats(self) -> None:
+        """Test default stats values."""
+        stats = CircuitStats()
+        assert stats.failures == 0
+        assert stats.successes == 0
+        assert stats.last_failure_time is None
+        assert stats.last_success_time is None
+        assert stats.half_open_calls == 0
+
+
+class TestCircuitBreaker:
+    """Tests for CircuitBreaker class."""
+
+    def test_initial_state(self) -> None:
+        """Test circuit breaker initial state."""
+        cb = CircuitBreaker(name="test", failure_threshold=5)
+
+        assert cb.name == "test"
+        assert cb.state == CircuitState.CLOSED
+        assert cb.failure_threshold == 5
+        assert cb.is_available() is True
+
+    def test_state_remains_closed_below_threshold(self) -> None:
+        """Test circuit stays closed below failure threshold."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+
+        # Record 2 failures (below threshold)
+        asyncio.run(cb.record_failure())
+        asyncio.run(cb.record_failure())
+
+        assert cb.state == CircuitState.CLOSED
+        assert cb.stats.failures == 2
+        assert cb.is_available() is True
+
+    def test_state_opens_at_threshold(self) -> None:
+        """Test circuit opens at failure threshold."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+
+        # Record 3 failures (at threshold)
+        asyncio.run(cb.record_failure())
+        asyncio.run(cb.record_failure())
+        asyncio.run(cb.record_failure())
+
+        assert cb.state == CircuitState.OPEN
+        assert cb.is_available() is False
+
+    def test_success_resets_in_closed(self) -> None:
+        """Test success in closed state records properly."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+
+        asyncio.run(cb.record_failure())
+        asyncio.run(cb.record_success())
+
+        assert cb.state == CircuitState.CLOSED
+        assert cb.stats.successes == 1
+        assert cb.stats.last_success_time is not None
+
+    def test_half_open_transition(self) -> None:
+        """Test transition to half-open after recovery timeout."""
+        cb = CircuitBreaker(
+            name="test",
+            failure_threshold=1,
+            recovery_timeout=1,  # 1 second
+        )
+
+        # Open the circuit
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+        # Wait for recovery timeout
+        time.sleep(1.1)
+
+        # State should transition to half-open
+        assert cb.state == CircuitState.HALF_OPEN
+        assert cb.is_available() is True
+
+    def test_half_open_success_closes(self) -> None:
+        """Test success in half-open closes circuit."""
+        cb = CircuitBreaker(
+            name="test",
+            failure_threshold=1,
+            recovery_timeout=0,  # Immediate recovery for testing
+        )
+
+        # Open and transition to half-open
+        asyncio.run(cb.record_failure())
+        time.sleep(0.1)
+        _ = cb.state  # Trigger state check
+
+        assert cb.state == CircuitState.HALF_OPEN
+
+        # Success should close
+        asyncio.run(cb.record_success())
+        assert cb.state == CircuitState.CLOSED
+
+    def test_half_open_failure_reopens(self) -> None:
+        """Test failure in half-open reopens circuit."""
+        cb = CircuitBreaker(
+            name="test",
+            failure_threshold=1,
+            recovery_timeout=0.05,  # Small but non-zero for reliable timing
+        )
+
+        # Open and transition to half-open
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+        # Wait for recovery timeout
+        time.sleep(0.1)
+        assert cb.state == CircuitState.HALF_OPEN
+
+        # Failure should reopen
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+    def test_half_open_call_limit(self) -> None:
+        """Test half-open call limit."""
+        cb = CircuitBreaker(
+            name="test",
+            failure_threshold=1,
+            recovery_timeout=0,
+            half_open_max_calls=2,
+        )
+
+        # Open and transition to half-open
+        asyncio.run(cb.record_failure())
+        time.sleep(0.1)
+        _ = cb.state
+
+        assert cb.is_available() is True
+
+        # Simulate calls in half-open
+        cb._stats.half_open_calls = 1
+        assert cb.is_available() is True
+
+        cb._stats.half_open_calls = 2
+        assert cb.is_available() is False
+
+    def test_time_until_recovery(self) -> None:
+        """Test time until recovery calculation."""
+        cb = CircuitBreaker(
+            name="test",
+            failure_threshold=1,
+            recovery_timeout=60,
+        )
+
+        # Closed circuit has no recovery time
+        assert cb.time_until_recovery() is None
+
+        # Open circuit
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+        # Should have recovery time
+        remaining = cb.time_until_recovery()
+        assert remaining is not None
+        assert 0 <= remaining <= 60
+
+    def test_execute_success(self) -> None:
+        """Test execute with successful function."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+
+        async def success_func() -> str:
+            return "success"
+
+        result = asyncio.run(cb.execute(success_func))
+        assert result == "success"
+        assert cb.stats.successes == 1
+
+    def test_execute_failure(self) -> None:
+        """Test execute with failing function."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+
+        async def fail_func() -> None:
+            raise ValueError("Error")
+
+        with pytest.raises(ValueError):
+            asyncio.run(cb.execute(fail_func))
+
+        assert cb.stats.failures == 1
+
+    def test_execute_when_open(self) -> None:
+        """Test execute raises when circuit is open."""
+        cb = CircuitBreaker(name="test", failure_threshold=1)
+
+        # Open the circuit
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+        async def success_func() -> str:
+            return "success"
+
+        with pytest.raises(CircuitOpenError) as exc_info:
+            asyncio.run(cb.execute(success_func))
+
+        assert exc_info.value.provider == "test"
+
+    def test_reset(self) -> None:
+        """Test circuit reset."""
+        cb = CircuitBreaker(name="test", failure_threshold=1)
+
+        # Open the circuit
+        asyncio.run(cb.record_failure())
+        assert cb.state == CircuitState.OPEN
+
+        # Reset
+        cb.reset()
+        assert cb.state == CircuitState.CLOSED
+        assert cb.stats.failures == 0
+        assert cb.stats.successes == 0
+
+    def test_to_dict(self) -> None:
+        """Test converting circuit to dict."""
+        cb = CircuitBreaker(name="test", failure_threshold=3)
+        asyncio.run(cb.record_failure())
+        asyncio.run(cb.record_success())
+
+        result = cb.to_dict()
+
+        assert result["name"] == "test"
+        assert result["state"] == "closed"
+        assert result["failures"] == 1
+        assert result["successes"] == 1
+        assert result["is_available"] is True
+
+
+class TestCircuitBreakerRegistry:
+    """Tests for CircuitBreakerRegistry class."""
+
+    def test_get_circuit_creates_new(self) -> None:
+        """Test getting a new circuit."""
+        settings = Settings(circuit_failure_threshold=5)
+        registry = CircuitBreakerRegistry(settings=settings)
+
+        circuit = asyncio.run(registry.get_circuit("anthropic"))
+
+        assert circuit.name == "anthropic"
+        assert circuit.failure_threshold == 5
+
+    def test_get_circuit_returns_same(self) -> None:
+        """Test getting same circuit twice."""
+        registry = CircuitBreakerRegistry()
+
+        circuit1 = asyncio.run(registry.get_circuit("openai"))
+        circuit2 = asyncio.run(registry.get_circuit("openai"))
+
+        assert circuit1 is circuit2
+
+    def test_get_circuit_sync(self) -> None:
+        """Test sync circuit getter."""
+        registry = CircuitBreakerRegistry()
+
+        circuit = registry.get_circuit_sync("google")
+        assert circuit.name == "google"
+
+    def test_is_available(self) -> None:
+        """Test checking if circuit is available."""
+        registry = CircuitBreakerRegistry()
+
+        assert asyncio.run(registry.is_available("test")) is True
+
+        # Open the circuit
+        circuit = asyncio.run(registry.get_circuit("test"))
+        for _ in range(5):
+            asyncio.run(circuit.record_failure())
+
+        assert asyncio.run(registry.is_available("test")) is False
+
+    def test_record_success(self) -> None:
+        """Test recording success through registry."""
+        registry = CircuitBreakerRegistry()
+
+        asyncio.run(registry.record_success("test"))
+
+        circuit = asyncio.run(registry.get_circuit("test"))
+        assert circuit.stats.successes == 1
+
+    def test_record_failure(self) -> None:
+        """Test recording failure through registry."""
+        registry = CircuitBreakerRegistry()
+
+        asyncio.run(registry.record_failure("test"))
+
+        circuit = asyncio.run(registry.get_circuit("test"))
+        assert circuit.stats.failures == 1
+
+    def test_reset(self) -> None:
+        """Test resetting a specific circuit."""
+        registry = CircuitBreakerRegistry()
+
+        # Create and fail a circuit
+        asyncio.run(registry.record_failure("test"))
+        asyncio.run(registry.reset("test"))
+
+        circuit = asyncio.run(registry.get_circuit("test"))
+        assert circuit.stats.failures == 0
+
+    def test_reset_all(self) -> None:
+        """Test resetting all circuits."""
+        registry = CircuitBreakerRegistry()
+
+        # Create multiple circuits with failures
+        asyncio.run(registry.record_failure("circuit1"))
+        asyncio.run(registry.record_failure("circuit2"))
+
+        asyncio.run(registry.reset_all())
+
+        circuit1 = asyncio.run(registry.get_circuit("circuit1"))
+        circuit2 = asyncio.run(registry.get_circuit("circuit2"))
+        assert circuit1.stats.failures == 0
+        assert circuit2.stats.failures == 0
+
+    def test_get_all_states(self) -> None:
+        """Test getting all circuit states."""
+        registry = CircuitBreakerRegistry()
+
+        asyncio.run(registry.get_circuit("circuit1"))
+        asyncio.run(registry.get_circuit("circuit2"))
+
+        states = registry.get_all_states()
+
+        assert "circuit1" in states
+        assert "circuit2" in states
+        assert states["circuit1"]["state"] == "closed"
+
+    def test_get_open_circuits(self) -> None:
+        """Test getting open circuits."""
+        settings = Settings(circuit_failure_threshold=1)
+        registry = CircuitBreakerRegistry(settings=settings)
+
+        asyncio.run(registry.get_circuit("healthy"))
+        asyncio.run(registry.record_failure("failing"))
+
+        open_circuits = registry.get_open_circuits()
+        assert "failing" in open_circuits
+        assert "healthy" not in open_circuits
+
+    def test_get_available_circuits(self) -> None:
+        """Test getting available circuits."""
+        settings = Settings(circuit_failure_threshold=1)
+        registry = CircuitBreakerRegistry(settings=settings)
+
+        asyncio.run(registry.get_circuit("healthy"))
+        asyncio.run(registry.record_failure("failing"))
+
+        available = registry.get_available_circuits()
+        assert "healthy" in available
+        assert "failing" not in available
+
+
+class TestGlobalRegistry:
+    """Tests for global registry functions."""
+
+    def test_get_circuit_registry(self) -> None:
+        """Test getting global registry."""
+        reset_circuit_registry()
+        registry = get_circuit_registry()
+        assert isinstance(registry, CircuitBreakerRegistry)
+
+    def test_get_circuit_registry_singleton(self) -> None:
+        """Test registry is singleton."""
+        reset_circuit_registry()
+        registry1 = get_circuit_registry()
+        registry2 = get_circuit_registry()
+        assert registry1 is registry2
+
+    def test_reset_circuit_registry(self) -> None:
+        """Test resetting global registry."""
+        reset_circuit_registry()
+        registry1 = get_circuit_registry()
+        reset_circuit_registry()
+        registry2 = get_circuit_registry()
+        assert registry1 is not registry2
--- a/mcp-servers/llm-gateway/tests/test_models.py
+++ b/mcp-servers/llm-gateway/tests/test_models.py
@@ -0,0 +1,408 @@
+"""
+Tests for models module.
+"""
+
+from datetime import UTC, datetime
+
+import pytest
+
+from models import (
+    AGENT_TYPE_MODEL_PREFERENCES,
+    MODEL_CONFIGS,
+    MODEL_GROUPS,
+    ChatMessage,
+    CompletionRequest,
+    CompletionResponse,
+    CostRecord,
+    EmbeddingRequest,
+    ModelConfig,
+    ModelGroup,
+    ModelGroupConfig,
+    ModelGroupInfo,
+    ModelInfo,
+    Provider,
+    StreamChunk,
+    UsageReport,
+    UsageStats,
+)
+
+
+class TestModelGroup:
+    """Tests for ModelGroup enum."""
+
+    def test_model_group_values(self) -> None:
+        """Test model group enum values."""
+        assert ModelGroup.REASONING.value == "reasoning"
+        assert ModelGroup.CODE.value == "code"
+        assert ModelGroup.FAST.value == "fast"
+        assert ModelGroup.VISION.value == "vision"
+        assert ModelGroup.EMBEDDING.value == "embedding"
+        assert ModelGroup.COST_OPTIMIZED.value == "cost_optimized"
+        assert ModelGroup.SELF_HOSTED.value == "self_hosted"
+
+    def test_model_group_from_string(self) -> None:
+        """Test creating ModelGroup from string."""
+        assert ModelGroup("reasoning") == ModelGroup.REASONING
+        assert ModelGroup("code") == ModelGroup.CODE
+        assert ModelGroup("fast") == ModelGroup.FAST
+
+    def test_model_group_invalid(self) -> None:
+        """Test invalid model group value."""
+        with pytest.raises(ValueError):
+            ModelGroup("invalid_group")
+
+
+class TestProvider:
+    """Tests for Provider enum."""
+
+    def test_provider_values(self) -> None:
+        """Test provider enum values."""
+        assert Provider.ANTHROPIC.value == "anthropic"
+        assert Provider.OPENAI.value == "openai"
+        assert Provider.GOOGLE.value == "google"
+        assert Provider.ALIBABA.value == "alibaba"
+        assert Provider.DEEPSEEK.value == "deepseek"
+
+
+class TestModelConfig:
+    """Tests for ModelConfig dataclass."""
+
+    def test_model_config_creation(self) -> None:
+        """Test creating a ModelConfig."""
+        config = ModelConfig(
+            name="test-model",
+            litellm_name="provider/test-model",
+            provider=Provider.ANTHROPIC,
+            cost_per_1m_input=10.0,
+            cost_per_1m_output=30.0,
+            context_window=100000,
+            max_output_tokens=4096,
+            supports_vision=True,
+        )
+
+        assert config.name == "test-model"
+        assert config.provider == Provider.ANTHROPIC
+        assert config.cost_per_1m_input == 10.0
+        assert config.supports_vision is True
+        assert config.supports_streaming is True  # default
+
+    def test_model_configs_exist(self) -> None:
+        """Test that model configs are defined."""
+        assert len(MODEL_CONFIGS) > 0
+        assert "claude-opus-4" in MODEL_CONFIGS
+        assert "gpt-4.1" in MODEL_CONFIGS
+        assert "gemini-2.5-pro" in MODEL_CONFIGS
+
+
+class TestModelGroupConfig:
+    """Tests for ModelGroupConfig dataclass."""
+
+    def test_model_group_config_creation(self) -> None:
+        """Test creating a ModelGroupConfig."""
+        config = ModelGroupConfig(
+            primary="model-a",
+            fallbacks=["model-b", "model-c"],
+            description="Test group",
+        )
+
+        assert config.primary == "model-a"
+        assert config.fallbacks == ["model-b", "model-c"]
+        assert config.description == "Test group"
+
+    def test_get_all_models(self) -> None:
+        """Test getting all models in order."""
+        config = ModelGroupConfig(
+            primary="model-a",
+            fallbacks=["model-b", "model-c"],
+            description="Test group",
+        )
+
+        models = config.get_all_models()
+        assert models == ["model-a", "model-b", "model-c"]
+
+    def test_model_groups_exist(self) -> None:
+        """Test that model groups are defined."""
+        assert len(MODEL_GROUPS) > 0
+        assert ModelGroup.REASONING in MODEL_GROUPS
+        assert ModelGroup.CODE in MODEL_GROUPS
+        assert ModelGroup.FAST in MODEL_GROUPS
+
+
+class TestAgentTypePreferences:
+    """Tests for agent type model preferences."""
+
+    def test_agent_preferences_exist(self) -> None:
+        """Test that agent preferences are defined."""
+        assert len(AGENT_TYPE_MODEL_PREFERENCES) > 0
+        assert "product_owner" in AGENT_TYPE_MODEL_PREFERENCES
+        assert "software_engineer" in AGENT_TYPE_MODEL_PREFERENCES
+
+    def test_agent_preference_values(self) -> None:
+        """Test agent preference values."""
+        assert AGENT_TYPE_MODEL_PREFERENCES["product_owner"] == ModelGroup.REASONING
+        assert AGENT_TYPE_MODEL_PREFERENCES["software_engineer"] == ModelGroup.CODE
+        assert AGENT_TYPE_MODEL_PREFERENCES["devops_engineer"] == ModelGroup.FAST
+
+
+class TestChatMessage:
+    """Tests for ChatMessage model."""
+
+    def test_chat_message_creation(self) -> None:
+        """Test creating a ChatMessage."""
+        msg = ChatMessage(role="user", content="Hello")
+        assert msg.role == "user"
+        assert msg.content == "Hello"
+        assert msg.name is None
+        assert msg.tool_calls is None
+
+    def test_chat_message_with_optional(self) -> None:
+        """Test ChatMessage with optional fields."""
+        msg = ChatMessage(
+            role="assistant",
+            content="Response",
+            name="assistant_1",
+            tool_calls=[{"id": "call_1", "function": {"name": "test"}}],
+        )
+        assert msg.name == "assistant_1"
+        assert msg.tool_calls is not None
+
+    def test_chat_message_list_content(self) -> None:
+        """Test ChatMessage with list content (for images)."""
+        msg = ChatMessage(
+            role="user",
+            content=[
+                {"type": "text", "text": "What's in this image?"},
+                {"type": "image_url", "image_url": {"url": "http://example.com/img.jpg"}},
+            ],
+        )
+        assert isinstance(msg.content, list)
+
+
+class TestCompletionRequest:
+    """Tests for CompletionRequest model."""
+
+    def test_completion_request_minimal(self) -> None:
+        """Test minimal CompletionRequest."""
+        req = CompletionRequest(
+            project_id="proj-123",
+            agent_id="agent-456",
+            messages=[ChatMessage(role="user", content="Hi")],
+        )
+
+        assert req.project_id == "proj-123"
+        assert req.agent_id == "agent-456"
+        assert len(req.messages) == 1
+        assert req.model_group == ModelGroup.REASONING  # default
+        assert req.max_tokens == 4096  # default
+        assert req.temperature == 0.7  # default
+
+    def test_completion_request_full(self) -> None:
+        """Test full CompletionRequest."""
+        req = CompletionRequest(
+            project_id="proj-123",
+            agent_id="agent-456",
+            messages=[ChatMessage(role="user", content="Hi")],
+            model_group=ModelGroup.CODE,
+            model_override="claude-sonnet-4",
+            max_tokens=8192,
+            temperature=0.5,
+            stream=True,
+            session_id="session-789",
+            metadata={"key": "value"},
+        )
+
+        assert req.model_group == ModelGroup.CODE
+        assert req.model_override == "claude-sonnet-4"
+        assert req.max_tokens == 8192
+        assert req.stream is True
+
+    def test_completion_request_validation(self) -> None:
+        """Test CompletionRequest validation."""
+        with pytest.raises(ValueError):
+            CompletionRequest(
+                project_id="proj-123",
+                agent_id="agent-456",
+                messages=[ChatMessage(role="user", content="Hi")],
+                max_tokens=0,  # Invalid
+            )
+
+        with pytest.raises(ValueError):
+            CompletionRequest(
+                project_id="proj-123",
+                agent_id="agent-456",
+                messages=[ChatMessage(role="user", content="Hi")],
+                temperature=-0.1,  # Invalid
+            )
+
+
+class TestUsageStats:
+    """Tests for UsageStats model."""
+
+    def test_usage_stats_default(self) -> None:
+        """Test default UsageStats."""
+        stats = UsageStats()
+        assert stats.prompt_tokens == 0
+        assert stats.completion_tokens == 0
+        assert stats.total_tokens == 0
+        assert stats.cost_usd == 0.0
+
+    def test_usage_stats_custom(self) -> None:
+        """Test custom UsageStats."""
+        stats = UsageStats(
+            prompt_tokens=100,
+            completion_tokens=50,
+            total_tokens=150,
+            cost_usd=0.001,
+        )
+        assert stats.prompt_tokens == 100
+        assert stats.total_tokens == 150
+
+    def test_usage_stats_from_response(self) -> None:
+        """Test creating UsageStats from response."""
+        config = MODEL_CONFIGS["claude-opus-4"]
+        stats = UsageStats.from_response(
+            prompt_tokens=1000,
+            completion_tokens=500,
+            model_config=config,
+        )
+
+        assert stats.prompt_tokens == 1000
+        assert stats.completion_tokens == 500
+        assert stats.total_tokens == 1500
+        # 1000/1M * 15 + 500/1M * 75 = 0.015 + 0.0375 = 0.0525
+        assert stats.cost_usd == pytest.approx(0.0525, rel=0.01)
+
+
+class TestCompletionResponse:
+    """Tests for CompletionResponse model."""
+
+    def test_completion_response_creation(self) -> None:
+        """Test creating a CompletionResponse."""
+        response = CompletionResponse(
+            id="resp-123",
+            model="claude-opus-4",
+            provider="anthropic",
+            content="Hello, world!",
+        )
+
+        assert response.id == "resp-123"
+        assert response.model == "claude-opus-4"
+        assert response.provider == "anthropic"
+        assert response.content == "Hello, world!"
+        assert response.finish_reason == "stop"
+
+
+class TestStreamChunk:
+    """Tests for StreamChunk model."""
+
+    def test_stream_chunk_creation(self) -> None:
+        """Test creating a StreamChunk."""
+        chunk = StreamChunk(id="chunk-1", delta="Hello")
+        assert chunk.id == "chunk-1"
+        assert chunk.delta == "Hello"
+        assert chunk.finish_reason is None
+
+    def test_stream_chunk_final(self) -> None:
+        """Test final StreamChunk."""
+        chunk = StreamChunk(
+            id="chunk-last",
+            delta="",
+            finish_reason="stop",
+            usage=UsageStats(prompt_tokens=10, completion_tokens=5, total_tokens=15),
+        )
+        assert chunk.finish_reason == "stop"
+        assert chunk.usage is not None
+
+
+class TestEmbeddingRequest:
+    """Tests for EmbeddingRequest model."""
+
+    def test_embedding_request_creation(self) -> None:
+        """Test creating an EmbeddingRequest."""
+        req = EmbeddingRequest(
+            project_id="proj-123",
+            agent_id="agent-456",
+            texts=["Hello", "World"],
+        )
+
+        assert req.project_id == "proj-123"
+        assert len(req.texts) == 2
+        assert req.model == "text-embedding-3-large"  # default
+
+    def test_embedding_request_validation(self) -> None:
+        """Test EmbeddingRequest validation."""
+        with pytest.raises(ValueError):
+            EmbeddingRequest(
+                project_id="proj-123",
+                agent_id="agent-456",
+                texts=[],  # Invalid - must have at least 1
+            )
+
+
+class TestCostRecord:
+    """Tests for CostRecord dataclass."""
+
+    def test_cost_record_creation(self) -> None:
+        """Test creating a CostRecord."""
+        record = CostRecord(
+            project_id="proj-123",
+            agent_id="agent-456",
+            model="claude-opus-4",
+            prompt_tokens=100,
+            completion_tokens=50,
+            cost_usd=0.01,
+        )
+
+        assert record.project_id == "proj-123"
+        assert record.cost_usd == 0.01
+        assert record.timestamp is not None
+
+
+class TestUsageReport:
+    """Tests for UsageReport model."""
+
+    def test_usage_report_creation(self) -> None:
+        """Test creating a UsageReport."""
+        now = datetime.now(UTC)
+        report = UsageReport(
+            entity_id="proj-123",
+            entity_type="project",
+            period="day",
+            period_start=now,
+            period_end=now,
+        )
+
+        assert report.entity_id == "proj-123"
+        assert report.entity_type == "project"
+        assert report.total_requests == 0
+        assert report.total_cost_usd == 0.0
+
+
+class TestModelInfo:
+    """Tests for ModelInfo model."""
+
+    def test_model_info_from_config(self) -> None:
+        """Test creating ModelInfo from ModelConfig."""
+        config = MODEL_CONFIGS["claude-opus-4"]
+        info = ModelInfo.from_config(config, available=True)
+
+        assert info.name == "claude-opus-4"
+        assert info.provider == "anthropic"
+        assert info.available is True
+        assert info.supports_vision is True
+
+
+class TestModelGroupInfo:
+    """Tests for ModelGroupInfo model."""
+
+    def test_model_group_info_creation(self) -> None:
+        """Test creating ModelGroupInfo."""
+        info = ModelGroupInfo(
+            name="reasoning",
+            description="Complex analysis",
+            primary_model="claude-opus-4",
+            fallback_models=["gpt-4.1"],
+        )
+
+        assert info.name == "reasoning"
+        assert len(info.fallback_models) == 1
--- a/mcp-servers/llm-gateway/tests/test_providers.py
+++ b/mcp-servers/llm-gateway/tests/test_providers.py
@@ -0,0 +1,308 @@
+"""
+Tests for providers module.
+"""
+
+import os
+from unittest.mock import patch
+
+import pytest
+
+from config import Settings
+from models import MODEL_CONFIGS, ModelGroup, Provider
+from providers import (
+    LLMProvider,
+    build_fallback_config,
+    build_model_list,
+    configure_litellm,
+    get_available_model_groups,
+    get_available_models,
+    get_provider,
+    reset_provider,
+)
+
+
+@pytest.fixture
+def full_settings() -> Settings:
+    """Settings with all providers configured."""
+    return Settings(
+        anthropic_api_key="test-anthropic-key",
+        openai_api_key="test-openai-key",
+        google_api_key="test-google-key",
+        alibaba_api_key="test-alibaba-key",
+        deepseek_api_key="test-deepseek-key",
+        litellm_timeout=60,
+        litellm_cache_enabled=False,
+    )
+
+
+@pytest.fixture
+def partial_settings() -> Settings:
+    """Settings with only some providers configured."""
+    return Settings(
+        anthropic_api_key="test-anthropic-key",
+        openai_api_key=None,
+        google_api_key=None,
+        alibaba_api_key=None,
+        deepseek_api_key=None,
+    )
+
+
+@pytest.fixture
+def empty_settings() -> Settings:
+    """Settings with no providers configured."""
+    return Settings(
+        anthropic_api_key=None,
+        openai_api_key=None,
+        google_api_key=None,
+        alibaba_api_key=None,
+        deepseek_api_key=None,
+    )
+
+
+class TestConfigureLiteLLM:
+    """Tests for configure_litellm function."""
+
+    def test_sets_api_keys(self, full_settings: Settings) -> None:
+        """Test that API keys are set in environment."""
+        with patch.dict(os.environ, {}, clear=True):
+            configure_litellm(full_settings)
+
+            assert os.environ.get("ANTHROPIC_API_KEY") == "test-anthropic-key"
+            assert os.environ.get("OPENAI_API_KEY") == "test-openai-key"
+            assert os.environ.get("GEMINI_API_KEY") == "test-google-key"
+
+    def test_skips_none_keys(self, partial_settings: Settings) -> None:
+        """Test that None keys are not set."""
+        with patch.dict(os.environ, {}, clear=True):
+            configure_litellm(partial_settings)
+
+            assert os.environ.get("ANTHROPIC_API_KEY") == "test-anthropic-key"
+            assert "OPENAI_API_KEY" not in os.environ
+
+
+class TestBuildModelList:
+    """Tests for build_model_list function."""
+
+    def test_build_with_all_providers(self, full_settings: Settings) -> None:
+        """Test building model list with all providers."""
+        model_list = build_model_list(full_settings)
+
+        assert len(model_list) > 0
+
+        # Check structure
+        for entry in model_list:
+            assert "model_name" in entry
+            assert "litellm_params" in entry
+            assert "model" in entry["litellm_params"]
+            assert "timeout" in entry["litellm_params"]
+
+    def test_build_with_partial_providers(self, partial_settings: Settings) -> None:
+        """Test building model list with partial providers."""
+        model_list = build_model_list(partial_settings)
+
+        # Should only include Anthropic models
+        providers = set()
+        for entry in model_list:
+            model_name = entry["model_name"]
+            config = MODEL_CONFIGS.get(model_name)
+            if config:
+                providers.add(config.provider)
+
+        assert Provider.ANTHROPIC in providers
+        assert Provider.OPENAI not in providers
+
+    def test_build_with_no_providers(self, empty_settings: Settings) -> None:
+        """Test building model list with no providers."""
+        model_list = build_model_list(empty_settings)
+
+        assert len(model_list) == 0
+
+    def test_build_includes_timeout(self, full_settings: Settings) -> None:
+        """Test that model entries include timeout."""
+        model_list = build_model_list(full_settings)
+
+        for entry in model_list:
+            assert entry["litellm_params"]["timeout"] == 60
+
+
+class TestBuildFallbackConfig:
+    """Tests for build_fallback_config function."""
+
+    def test_build_fallbacks_full(self, full_settings: Settings) -> None:
+        """Test building fallback config with all providers."""
+        fallbacks = build_fallback_config(full_settings)
+
+        assert len(fallbacks) > 0
+
+        # Primary models should have fallbacks
+        for _primary, chain in fallbacks.items():
+            assert isinstance(chain, list)
+            assert len(chain) > 0
+
+    def test_build_fallbacks_partial(self, partial_settings: Settings) -> None:
+        """Test building fallback config with partial providers."""
+        fallbacks = build_fallback_config(partial_settings)
+
+        # With only Anthropic, there should be no fallbacks
+        # (fallbacks require at least 2 available models)
+        for primary, chain in fallbacks.items():
+            # All models in chain should be from Anthropic
+            for model in [primary] + chain:
+                config = MODEL_CONFIGS.get(model)
+                if config:
+                    assert config.provider == Provider.ANTHROPIC
+
+
+class TestGetAvailableModels:
+    """Tests for get_available_models function."""
+
+    def test_get_available_full(self, full_settings: Settings) -> None:
+        """Test getting available models with all providers."""
+        models = get_available_models(full_settings)
+
+        assert len(models) > 0
+        assert "claude-opus-4" in models
+        assert "gpt-4.1" in models
+
+    def test_get_available_partial(self, partial_settings: Settings) -> None:
+        """Test getting available models with partial providers."""
+        models = get_available_models(partial_settings)
+
+        assert "claude-opus-4" in models
+        assert "gpt-4.1" not in models
+
+    def test_get_available_empty(self, empty_settings: Settings) -> None:
+        """Test getting available models with no providers."""
+        models = get_available_models(empty_settings)
+
+        assert len(models) == 0
+
+
+class TestGetAvailableModelGroups:
+    """Tests for get_available_model_groups function."""
+
+    def test_get_groups_full(self, full_settings: Settings) -> None:
+        """Test getting groups with all providers."""
+        groups = get_available_model_groups(full_settings)
+
+        assert len(groups) == len(ModelGroup)
+        assert ModelGroup.REASONING in groups
+        assert len(groups[ModelGroup.REASONING]) > 0
+
+    def test_get_groups_partial(self, partial_settings: Settings) -> None:
+        """Test getting groups with partial providers."""
+        groups = get_available_model_groups(partial_settings)
+
+        # Only Anthropic models should be available
+        for _group, models in groups.items():
+            for model in models:
+                config = MODEL_CONFIGS.get(model)
+                if config:
+                    assert config.provider == Provider.ANTHROPIC
+
+
+class TestLLMProvider:
+    """Tests for LLMProvider class."""
+
+    def test_initialization(self, full_settings: Settings) -> None:
+        """Test provider initialization."""
+        provider = LLMProvider(settings=full_settings)
+
+        assert provider._initialized is False
+        assert provider._router is None
+
+    def test_initialize(self, full_settings: Settings) -> None:
+        """Test provider initialize."""
+        with patch("providers.Router") as mock_router:
+            provider = LLMProvider(settings=full_settings)
+            provider.initialize()
+
+            assert provider._initialized is True
+            mock_router.assert_called_once()
+
+    def test_initialize_idempotent(self, full_settings: Settings) -> None:
+        """Test that initialize is idempotent."""
+        with patch("providers.Router") as mock_router:
+            provider = LLMProvider(settings=full_settings)
+            provider.initialize()
+            provider.initialize()
+
+            # Should only be called once
+            assert mock_router.call_count == 1
+
+    def test_initialize_no_providers(self, empty_settings: Settings) -> None:
+        """Test initialization with no providers."""
+        provider = LLMProvider(settings=empty_settings)
+        provider.initialize()
+
+        assert provider._initialized is True
+        assert provider._router is None
+
+    def test_router_property(self, full_settings: Settings) -> None:
+        """Test router property triggers initialization."""
+        with patch("providers.Router"):
+            provider = LLMProvider(settings=full_settings)
+            _ = provider.router
+
+            assert provider._initialized is True
+
+    def test_is_available(self, full_settings: Settings) -> None:
+        """Test is_available property."""
+        with patch("providers.Router"):
+            provider = LLMProvider(settings=full_settings)
+            assert provider.is_available is True
+
+    def test_is_not_available(self, empty_settings: Settings) -> None:
+        """Test is_available when no providers."""
+        provider = LLMProvider(settings=empty_settings)
+        assert provider.is_available is False
+
+    def test_get_model_config(self, full_settings: Settings) -> None:
+        """Test getting model config."""
+        provider = LLMProvider(settings=full_settings)
+
+        config = provider.get_model_config("claude-opus-4")
+        assert config is not None
+        assert config.name == "claude-opus-4"
+
+        assert provider.get_model_config("nonexistent") is None
+
+    def test_get_available_models(self, full_settings: Settings) -> None:
+        """Test getting available models."""
+        provider = LLMProvider(settings=full_settings)
+        models = provider.get_available_models()
+
+        assert "claude-opus-4" in models
+        assert "gpt-4.1" in models
+
+    def test_is_model_available(self, full_settings: Settings) -> None:
+        """Test checking model availability."""
+        provider = LLMProvider(settings=full_settings)
+
+        assert provider.is_model_available("claude-opus-4") is True
+        assert provider.is_model_available("nonexistent") is False
+
+
+class TestGlobalProvider:
+    """Tests for global provider functions."""
+
+    def test_get_provider(self) -> None:
+        """Test getting global provider."""
+        reset_provider()
+        provider = get_provider()
+        assert isinstance(provider, LLMProvider)
+
+    def test_get_provider_singleton(self) -> None:
+        """Test provider is singleton."""
+        reset_provider()
+        provider1 = get_provider()
+        provider2 = get_provider()
+        assert provider1 is provider2
+
+    def test_reset_provider(self) -> None:
+        """Test resetting global provider."""
+        reset_provider()
+        provider1 = get_provider()
+        reset_provider()
+        provider2 = get_provider()
+        assert provider1 is not provider2
--- a/mcp-servers/llm-gateway/tests/test_routing.py
+++ b/mcp-servers/llm-gateway/tests/test_routing.py
@@ -0,0 +1,243 @@
+"""
+Tests for routing module.
+"""
+
+import asyncio
+
+import pytest
+
+from config import Settings
+from exceptions import (
+    AllProvidersFailedError,
+    InvalidModelError,
+    InvalidModelGroupError,
+    ModelNotAvailableError,
+)
+from failover import CircuitBreakerRegistry, reset_circuit_registry
+from models import ModelGroup
+from providers import reset_provider
+from routing import ModelRouter, get_model_router, reset_model_router
+
+
+@pytest.fixture
+def router_settings() -> Settings:
+    """Settings for routing tests."""
+    return Settings(
+        anthropic_api_key="test-key",
+        openai_api_key="test-key",
+        google_api_key="test-key",
+    )
+
+
+@pytest.fixture
+def router(router_settings: Settings) -> ModelRouter:
+    """Create model router for testing."""
+    reset_circuit_registry()
+    reset_model_router()
+    reset_provider()
+    registry = CircuitBreakerRegistry(settings=router_settings)
+    return ModelRouter(settings=router_settings, circuit_registry=registry)
+
+
+class TestModelRouter:
+    """Tests for ModelRouter class."""
+
+    def test_parse_model_group_valid(self, router: ModelRouter) -> None:
+        """Test parsing valid model groups."""
+        assert router.parse_model_group("reasoning") == ModelGroup.REASONING
+        assert router.parse_model_group("code") == ModelGroup.CODE
+        assert router.parse_model_group("fast") == ModelGroup.FAST
+        assert router.parse_model_group("REASONING") == ModelGroup.REASONING
+
+    def test_parse_model_group_aliases(self, router: ModelRouter) -> None:
+        """Test parsing model group aliases."""
+        assert router.parse_model_group("high-reasoning") == ModelGroup.REASONING
+        assert router.parse_model_group("high_reasoning") == ModelGroup.REASONING
+        assert router.parse_model_group("code-generation") == ModelGroup.CODE
+        assert router.parse_model_group("fast-response") == ModelGroup.FAST
+
+    def test_parse_model_group_invalid(self, router: ModelRouter) -> None:
+        """Test parsing invalid model group."""
+        with pytest.raises(InvalidModelGroupError) as exc_info:
+            router.parse_model_group("invalid_group")
+
+        assert exc_info.value.model_group == "invalid_group"
+        assert exc_info.value.available_groups is not None
+
+    def test_get_model_config_valid(self, router: ModelRouter) -> None:
+        """Test getting valid model config."""
+        config = router.get_model_config("claude-opus-4")
+        assert config.name == "claude-opus-4"
+        assert config.provider.value == "anthropic"
+
+    def test_get_model_config_invalid(self, router: ModelRouter) -> None:
+        """Test getting invalid model config."""
+        with pytest.raises(InvalidModelError) as exc_info:
+            router.get_model_config("nonexistent-model")
+
+        assert exc_info.value.model == "nonexistent-model"
+
+    def test_get_preferred_group_for_agent(self, router: ModelRouter) -> None:
+        """Test getting preferred group for agent types."""
+        assert router.get_preferred_group_for_agent("product_owner") == ModelGroup.REASONING
+        assert router.get_preferred_group_for_agent("software_engineer") == ModelGroup.CODE
+        assert router.get_preferred_group_for_agent("devops_engineer") == ModelGroup.FAST
+
+    def test_get_preferred_group_unknown_agent(self, router: ModelRouter) -> None:
+        """Test getting preferred group for unknown agent."""
+        # Should default to REASONING
+        assert router.get_preferred_group_for_agent("unknown_type") == ModelGroup.REASONING
+
+    def test_select_model_by_group(self, router: ModelRouter) -> None:
+        """Test selecting model by group."""
+        model_name, config = asyncio.run(
+            router.select_model(model_group=ModelGroup.REASONING)
+        )
+
+        assert model_name == "claude-opus-4"
+        assert config.provider.value == "anthropic"
+
+    def test_select_model_by_group_string(self, router: ModelRouter) -> None:
+        """Test selecting model by group string."""
+        model_name, config = asyncio.run(
+            router.select_model(model_group="code")
+        )
+
+        assert model_name == "claude-sonnet-4"
+
+    def test_select_model_with_override(self, router: ModelRouter) -> None:
+        """Test selecting specific model override."""
+        model_name, config = asyncio.run(
+            router.select_model(
+                model_group="reasoning",
+                model_override="gpt-4.1",
+            )
+        )
+
+        assert model_name == "gpt-4.1"
+        assert config.provider.value == "openai"
+
+    def test_select_model_override_invalid(self, router: ModelRouter) -> None:
+        """Test selecting invalid model override."""
+        with pytest.raises(InvalidModelError):
+            asyncio.run(
+                router.select_model(
+                    model_group="reasoning",
+                    model_override="nonexistent-model",
+                )
+            )
+
+    def test_select_model_override_unavailable(self, router: ModelRouter) -> None:  # noqa: ARG002
+        """Test selecting unavailable model override."""
+        # Create router without Alibaba key
+        settings = Settings(
+            anthropic_api_key="test-key",
+            alibaba_api_key=None,
+        )
+        registry = CircuitBreakerRegistry(settings=settings)
+        limited_router = ModelRouter(settings=settings, circuit_registry=registry)
+
+        with pytest.raises(ModelNotAvailableError):
+            asyncio.run(
+                limited_router.select_model(
+                    model_group="reasoning",
+                    model_override="qwen-max",
+                )
+            )
+
+    def test_select_model_fallback_on_circuit_open(
+        self,
+        router: ModelRouter,
+    ) -> None:
+        """Test fallback when primary circuit is open."""
+        # Open circuit for anthropic
+        circuit = router._circuit_registry.get_circuit_sync("anthropic")
+        for _ in range(5):
+            asyncio.run(circuit.record_failure())
+
+        # Should fall back to OpenAI
+        model_name, config = asyncio.run(
+            router.select_model(model_group=ModelGroup.REASONING)
+        )
+
+        assert model_name == "gpt-4.1"
+        assert config.provider.value == "openai"
+
+    def test_select_model_all_unavailable(self) -> None:
+        """Test when all providers are unavailable."""
+        settings = Settings(
+            anthropic_api_key=None,
+            openai_api_key=None,
+            google_api_key=None,
+        )
+        registry = CircuitBreakerRegistry(settings=settings)
+        limited_router = ModelRouter(settings=settings, circuit_registry=registry)
+
+        with pytest.raises(AllProvidersFailedError) as exc_info:
+            asyncio.run(
+                limited_router.select_model(model_group=ModelGroup.REASONING)
+            )
+
+        assert exc_info.value.model_group == "reasoning"
+        assert len(exc_info.value.attempted_models) > 0
+
+    def test_get_available_models_for_group(self, router: ModelRouter) -> None:
+        """Test getting available models for a group."""
+        models = asyncio.run(
+            router.get_available_models_for_group(ModelGroup.REASONING)
+        )
+
+        assert len(models) > 0
+        # Should be (name, config, available) tuples
+        for name, config, _available in models:
+            assert isinstance(name, str)
+            assert config is not None
+
+    def test_get_available_models_for_group_string(self, router: ModelRouter) -> None:
+        """Test getting available models with string group."""
+        models = asyncio.run(
+            router.get_available_models_for_group("code")
+        )
+
+        assert len(models) > 0
+
+    def test_get_available_models_invalid_group(self, router: ModelRouter) -> None:
+        """Test getting models for invalid group."""
+        with pytest.raises(InvalidModelGroupError):
+            asyncio.run(
+                router.get_available_models_for_group("invalid")
+            )
+
+    def test_get_all_model_groups(self, router: ModelRouter) -> None:
+        """Test getting all model groups info."""
+        groups = router.get_all_model_groups()
+
+        assert len(groups) == len(ModelGroup)
+        assert "reasoning" in groups
+        assert "code" in groups
+        assert groups["reasoning"]["primary"] == "claude-opus-4"
+
+
+class TestGlobalRouter:
+    """Tests for global router functions."""
+
+    def test_get_model_router(self) -> None:
+        """Test getting global router."""
+        reset_model_router()
+        router = get_model_router()
+        assert isinstance(router, ModelRouter)
+
+    def test_get_model_router_singleton(self) -> None:
+        """Test router is singleton."""
+        reset_model_router()
+        router1 = get_model_router()
+        router2 = get_model_router()
+        assert router1 is router2
+
+    def test_reset_model_router(self) -> None:
+        """Test resetting global router."""
+        reset_model_router()
+        router1 = get_model_router()
+        reset_model_router()
+        router2 = get_model_router()
+        assert router1 is not router2
--- a/mcp-servers/llm-gateway/tests/test_server.py
+++ b/mcp-servers/llm-gateway/tests/test_server.py
@@ -0,0 +1,412 @@
+"""
+Tests for server module.
+"""
+
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from fastapi.testclient import TestClient
+
+from config import Settings
+from models import ModelGroup
+
+
+@pytest.fixture
+def test_settings() -> Settings:
+    """Test settings with mock API keys."""
+    return Settings(
+        anthropic_api_key="test-anthropic-key",
+        openai_api_key="test-openai-key",
+        google_api_key="test-google-key",
+        cost_tracking_enabled=False,  # Disable for most tests
+        litellm_cache_enabled=False,
+    )
+
+
+@pytest.fixture
+def test_client(test_settings: Settings) -> TestClient:
+    """Create test client with mocked dependencies."""
+    with (
+        patch("server.get_settings", return_value=test_settings),
+        patch("server.get_provider") as mock_provider,
+    ):
+        mock_provider.return_value = MagicMock()
+        mock_provider.return_value.is_available = True
+        mock_provider.return_value.router = MagicMock()
+        mock_provider.return_value.get_available_models.return_value = {}
+
+        from server import app
+        return TestClient(app)
+
+
+class TestHealthEndpoint:
+    """Tests for health check endpoint."""
+
+    def test_health_check(self, test_client: TestClient) -> None:
+        """Test health check returns healthy status."""
+        with patch("server.get_settings") as mock_settings:
+            mock_settings.return_value = Settings(
+                anthropic_api_key="test-key",
+            )
+            with patch("server.get_provider") as mock_provider:
+                mock_provider.return_value = MagicMock()
+                mock_provider.return_value.is_available = True
+
+                response = test_client.get("/health")
+
+                assert response.status_code == 200
+                data = response.json()
+                assert data["status"] == "healthy"
+                assert data["service"] == "llm-gateway"
+
+
+class TestToolDiscoveryEndpoint:
+    """Tests for tool discovery endpoint."""
+
+    def test_list_tools(self, test_client: TestClient) -> None:
+        """Test listing available tools."""
+        response = test_client.get("/mcp/tools")
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "tools" in data
+        assert len(data["tools"]) == 4  # 4 tools defined
+
+        tool_names = [t["name"] for t in data["tools"]]
+        assert "chat_completion" in tool_names
+        assert "list_models" in tool_names
+        assert "get_usage" in tool_names
+        assert "count_tokens" in tool_names
+
+    def test_tool_has_schema(self, test_client: TestClient) -> None:
+        """Test that tools have input schemas."""
+        response = test_client.get("/mcp/tools")
+        data = response.json()
+
+        for tool in data["tools"]:
+            assert "inputSchema" in tool
+            assert "type" in tool["inputSchema"]
+            assert tool["inputSchema"]["type"] == "object"
+
+
+class TestJSONRPCEndpoint:
+    """Tests for JSON-RPC endpoint."""
+
+    def test_invalid_jsonrpc_version(self, test_client: TestClient) -> None:
+        """Test invalid JSON-RPC version."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "1.0",  # Invalid
+                "method": "tools/list",
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "error" in data
+        assert data["error"]["code"] == -32600
+
+    def test_tools_list(self, test_client: TestClient) -> None:
+        """Test tools/list method."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "2.0",
+                "method": "tools/list",
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "result" in data
+        assert "tools" in data["result"]
+
+    def test_unknown_method(self, test_client: TestClient) -> None:
+        """Test unknown method."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "2.0",
+                "method": "unknown/method",
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "error" in data
+        assert data["error"]["code"] == -32601
+
+    def test_unknown_tool(self, test_client: TestClient) -> None:
+        """Test unknown tool."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "2.0",
+                "method": "tools/call",
+                "params": {
+                    "name": "unknown_tool",
+                    "arguments": {},
+                },
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "error" in data
+        assert "Unknown tool" in data["error"]["message"]
+
+
+class TestCountTokensTool:
+    """Tests for count_tokens tool."""
+
+    def test_count_tokens(self, test_client: TestClient) -> None:
+        """Test counting tokens."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "2.0",
+                "method": "tools/call",
+                "params": {
+                    "name": "count_tokens",
+                    "arguments": {
+                        "project_id": "proj-123",
+                        "agent_id": "agent-456",
+                        "text": "Hello, world!",
+                    },
+                },
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+        data = response.json()
+        assert "result" in data
+
+    def test_count_tokens_with_model(self, test_client: TestClient) -> None:
+        """Test counting tokens with specific model."""
+        response = test_client.post(
+            "/mcp",
+            json={
+                "jsonrpc": "2.0",
+                "method": "tools/call",
+                "params": {
+                    "name": "count_tokens",
+                    "arguments": {
+                        "project_id": "proj-123",
+                        "agent_id": "agent-456",
+                        "text": "Hello, world!",
+                        "model": "gpt-4",
+                    },
+                },
+                "id": 1,
+            },
+        )
+
+        assert response.status_code == 200
+
+
+class TestListModelsTool:
+    """Tests for list_models tool."""
+
+    def test_list_all_models(self, test_client: TestClient) -> None:
+        """Test listing all models."""
+        with patch("server.get_model_router") as mock_router:
+            mock_router.return_value = MagicMock()
+            mock_router.return_value.get_available_models_for_group = AsyncMock(
+                return_value=[]
+            )
+
+            response = test_client.post(
+                "/mcp",
+                json={
+                    "jsonrpc": "2.0",
+                    "method": "tools/call",
+                    "params": {
+                        "name": "list_models",
+                        "arguments": {
+                            "project_id": "proj-123",
+                            "agent_id": "agent-456",
+                        },
+                    },
+                    "id": 1,
+                },
+            )
+
+            assert response.status_code == 200
+
+    def test_list_models_by_group(self, test_client: TestClient) -> None:
+        """Test listing models by group."""
+        with patch("server.get_model_router") as mock_router:
+            mock_router.return_value = MagicMock()
+            mock_router.return_value.parse_model_group.return_value = ModelGroup.REASONING
+            mock_router.return_value.get_available_models_for_group = AsyncMock(
+                return_value=[]
+            )
+
+            response = test_client.post(
+                "/mcp",
+                json={
+                    "jsonrpc": "2.0",
+                    "method": "tools/call",
+                    "params": {
+                        "name": "list_models",
+                        "arguments": {
+                            "project_id": "proj-123",
+                            "agent_id": "agent-456",
+                            "model_group": "reasoning",
+                        },
+                    },
+                    "id": 1,
+                },
+            )
+
+            assert response.status_code == 200
+
+
+class TestGetUsageTool:
+    """Tests for get_usage tool."""
+
+    def test_get_usage(self, test_client: TestClient) -> None:
+        """Test getting usage."""
+        with patch("server.get_cost_tracker") as mock_tracker:
+            mock_report = MagicMock()
+            mock_report.total_requests = 10
+            mock_report.total_tokens = 1000
+            mock_report.total_cost_usd = 0.50
+            mock_report.by_model = {}
+            mock_report.period_start.isoformat.return_value = "2024-01-01T00:00:00"
+            mock_report.period_end.isoformat.return_value = "2024-01-02T00:00:00"
+
+            mock_tracker.return_value = MagicMock()
+            mock_tracker.return_value.get_project_usage = AsyncMock(
+                return_value=mock_report
+            )
+            mock_tracker.return_value.get_agent_usage = AsyncMock(
+                return_value=mock_report
+            )
+
+            response = test_client.post(
+                "/mcp",
+                json={
+                    "jsonrpc": "2.0",
+                    "method": "tools/call",
+                    "params": {
+                        "name": "get_usage",
+                        "arguments": {
+                            "project_id": "proj-123",
+                            "agent_id": "agent-456",
+                            "period": "day",
+                        },
+                    },
+                    "id": 1,
+                },
+            )
+
+            assert response.status_code == 200
+
+
+class TestChatCompletionTool:
+    """Tests for chat_completion tool."""
+
+    def test_chat_completion_streaming_not_supported(
+        self,
+        test_client: TestClient,
+    ) -> None:
+        """Test that streaming returns info message."""
+        with patch("server.get_model_router") as mock_router:
+            mock_router.return_value = MagicMock()
+            mock_router.return_value.select_model = AsyncMock(
+                return_value=("claude-opus-4", MagicMock())
+            )
+
+            with patch("server.get_cost_tracker") as mock_tracker:
+                mock_tracker.return_value = MagicMock()
+                mock_tracker.return_value.check_budget = AsyncMock(
+                    return_value=(True, 0.0, 100.0)
+                )
+
+                response = test_client.post(
+                    "/mcp",
+                    json={
+                        "jsonrpc": "2.0",
+                        "method": "tools/call",
+                        "params": {
+                            "name": "chat_completion",
+                            "arguments": {
+                                "project_id": "proj-123",
+                                "agent_id": "agent-456",
+                                "messages": [
+                                    {"role": "user", "content": "Hello"}
+                                ],
+                                "stream": True,
+                            },
+                        },
+                        "id": 1,
+                    },
+                )
+
+                assert response.status_code == 200
+
+    def test_chat_completion_success(self, test_client: TestClient) -> None:
+        """Test successful chat completion."""
+        mock_response = MagicMock()
+        mock_response.choices = [MagicMock()]
+        mock_response.choices[0].message.content = "Hello, world!"
+        mock_response.choices[0].finish_reason = "stop"
+        mock_response.usage = MagicMock()
+        mock_response.usage.prompt_tokens = 10
+        mock_response.usage.completion_tokens = 5
+
+        with patch("server.get_model_router") as mock_router:
+            mock_model_config = MagicMock()
+            mock_model_config.provider.value = "anthropic"
+            mock_router.return_value = MagicMock()
+            mock_router.return_value.select_model = AsyncMock(
+                return_value=("claude-opus-4", mock_model_config)
+            )
+
+            with patch("server.get_cost_tracker") as mock_tracker:
+                mock_tracker.return_value = MagicMock()
+                mock_tracker.return_value.check_budget = AsyncMock(
+                    return_value=(True, 0.0, 100.0)
+                )
+                mock_tracker.return_value.record_usage = AsyncMock()
+
+                with patch("server.get_provider") as mock_prov:
+                    mock_prov.return_value = MagicMock()
+                    mock_prov.return_value.router = MagicMock()
+                    mock_prov.return_value.router.acompletion = AsyncMock(
+                        return_value=mock_response
+                    )
+
+                    with patch("server.get_circuit_registry") as mock_reg:
+                        mock_circuit = MagicMock()
+                        mock_circuit.record_success = AsyncMock()
+                        mock_reg.return_value = MagicMock()
+                        mock_reg.return_value.get_circuit_sync.return_value = mock_circuit
+
+                        response = test_client.post(
+                            "/mcp",
+                            json={
+                                "jsonrpc": "2.0",
+                                "method": "tools/call",
+                                "params": {
+                                    "name": "chat_completion",
+                                    "arguments": {
+                                        "project_id": "proj-123",
+                                        "agent_id": "agent-456",
+                                        "messages": [
+                                            {"role": "user", "content": "Hello"}
+                                        ],
+                                    },
+                                },
+                                "id": 1,
+                            },
+                        )
+
+                        assert response.status_code == 200
--- a/mcp-servers/llm-gateway/tests/test_streaming.py
+++ b/mcp-servers/llm-gateway/tests/test_streaming.py
@@ -0,0 +1,312 @@
+"""
+Tests for streaming module.
+"""
+
+import asyncio
+import json
+from unittest.mock import MagicMock
+
+import pytest
+
+from models import StreamChunk, UsageStats
+from streaming import (
+    StreamAccumulator,
+    StreamBuffer,
+    format_sse_chunk,
+    format_sse_done,
+    format_sse_error,
+    stream_to_string,
+    wrap_litellm_stream,
+)
+
+
+class TestStreamAccumulator:
+    """Tests for StreamAccumulator class."""
+
+    def test_initial_state(self) -> None:
+        """Test initial accumulator state."""
+        acc = StreamAccumulator()
+
+        assert acc.request_id is not None
+        assert acc.content == ""
+        assert acc.chunks_received == 0
+        assert acc.prompt_tokens == 0
+        assert acc.completion_tokens == 0
+        assert acc.model is None
+        assert acc.finish_reason is None
+
+    def test_custom_request_id(self) -> None:
+        """Test accumulator with custom request ID."""
+        acc = StreamAccumulator(request_id="custom-id")
+        assert acc.request_id == "custom-id"
+
+    def test_add_chunk_text(self) -> None:
+        """Test adding text chunks."""
+        acc = StreamAccumulator()
+
+        acc.add_chunk("Hello")
+        acc.add_chunk(", ")
+        acc.add_chunk("world!")
+
+        assert acc.content == "Hello, world!"
+        assert acc.chunks_received == 3
+
+    def test_add_chunk_with_finish_reason(self) -> None:
+        """Test adding chunk with finish reason."""
+        acc = StreamAccumulator()
+
+        acc.add_chunk("Final", finish_reason="stop")
+
+        assert acc.finish_reason == "stop"
+
+    def test_add_chunk_with_model(self) -> None:
+        """Test adding chunk with model info."""
+        acc = StreamAccumulator()
+
+        acc.add_chunk("Text", model="claude-opus-4")
+
+        assert acc.model == "claude-opus-4"
+
+    def test_add_chunk_with_usage(self) -> None:
+        """Test adding chunk with usage stats."""
+        acc = StreamAccumulator()
+
+        acc.add_chunk(
+            "Text",
+            usage={"prompt_tokens": 10, "completion_tokens": 5},
+        )
+
+        assert acc.prompt_tokens == 10
+        assert acc.completion_tokens == 5
+        assert acc.total_tokens == 15
+
+    def test_start_and_finish(self) -> None:
+        """Test start and finish timing."""
+        acc = StreamAccumulator()
+
+        assert acc.duration_seconds is None
+
+        acc.start()
+        acc.finish()
+
+        assert acc.duration_seconds is not None
+        assert acc.duration_seconds >= 0
+
+    def test_get_usage_stats(self) -> None:
+        """Test getting usage stats."""
+        acc = StreamAccumulator()
+        acc.add_chunk("", usage={"prompt_tokens": 100, "completion_tokens": 50})
+
+        stats = acc.get_usage_stats(cost_usd=0.01)
+
+        assert stats.prompt_tokens == 100
+        assert stats.completion_tokens == 50
+        assert stats.total_tokens == 150
+        assert stats.cost_usd == 0.01
+
+
+class TestWrapLiteLLMStream:
+    """Tests for wrap_litellm_stream function."""
+
+    async def test_wrap_stream_basic(self) -> None:
+        """Test wrapping a basic stream."""
+        # Create mock stream chunks
+        async def mock_stream():
+            chunk1 = MagicMock()
+            chunk1.choices = [MagicMock()]
+            chunk1.choices[0].delta = MagicMock()
+            chunk1.choices[0].delta.content = "Hello"
+            chunk1.choices[0].finish_reason = None
+            chunk1.model = "test-model"
+            chunk1.usage = None
+            yield chunk1
+
+            chunk2 = MagicMock()
+            chunk2.choices = [MagicMock()]
+            chunk2.choices[0].delta = MagicMock()
+            chunk2.choices[0].delta.content = " World"
+            chunk2.choices[0].finish_reason = "stop"
+            chunk2.model = "test-model"
+            chunk2.usage = MagicMock()
+            chunk2.usage.prompt_tokens = 5
+            chunk2.usage.completion_tokens = 2
+            yield chunk2
+
+        accumulator = StreamAccumulator()
+        chunks = []
+
+        async for chunk in wrap_litellm_stream(mock_stream(), accumulator):
+            chunks.append(chunk)
+
+        assert len(chunks) == 2
+        assert chunks[0].delta == "Hello"
+        assert chunks[1].delta == " World"
+        assert chunks[1].finish_reason == "stop"
+        assert accumulator.content == "Hello World"
+
+    async def test_wrap_stream_without_accumulator(self) -> None:
+        """Test wrapping stream without accumulator."""
+        async def mock_stream():
+            chunk = MagicMock()
+            chunk.choices = [MagicMock()]
+            chunk.choices[0].delta = MagicMock()
+            chunk.choices[0].delta.content = "Test"
+            chunk.choices[0].finish_reason = None
+            chunk.model = None
+            chunk.usage = None
+            yield chunk
+
+        chunks = []
+        async for chunk in wrap_litellm_stream(mock_stream()):
+            chunks.append(chunk)
+
+        assert len(chunks) == 1
+
+
+class TestSSEFormatting:
+    """Tests for SSE formatting functions."""
+
+    def test_format_sse_chunk_basic(self) -> None:
+        """Test formatting basic chunk."""
+        chunk = StreamChunk(id="chunk-1", delta="Hello")
+        result = format_sse_chunk(chunk)
+
+        assert result.startswith("data: ")
+        assert result.endswith("\n\n")
+
+        # Parse the JSON
+        data = json.loads(result[6:-2])
+        assert data["id"] == "chunk-1"
+        assert data["delta"] == "Hello"
+
+    def test_format_sse_chunk_with_finish(self) -> None:
+        """Test formatting chunk with finish reason."""
+        chunk = StreamChunk(
+            id="chunk-2",
+            delta="",
+            finish_reason="stop",
+        )
+        result = format_sse_chunk(chunk)
+        data = json.loads(result[6:-2])
+
+        assert data["finish_reason"] == "stop"
+
+    def test_format_sse_chunk_with_usage(self) -> None:
+        """Test formatting chunk with usage."""
+        chunk = StreamChunk(
+            id="chunk-3",
+            delta="",
+            finish_reason="stop",
+            usage=UsageStats(
+                prompt_tokens=10,
+                completion_tokens=5,
+                total_tokens=15,
+                cost_usd=0.001,
+            ),
+        )
+        result = format_sse_chunk(chunk)
+        data = json.loads(result[6:-2])
+
+        assert "usage" in data
+        assert data["usage"]["prompt_tokens"] == 10
+
+    def test_format_sse_done(self) -> None:
+        """Test formatting done message."""
+        result = format_sse_done()
+        assert result == "data: [DONE]\n\n"
+
+    def test_format_sse_error(self) -> None:
+        """Test formatting error message."""
+        result = format_sse_error("Something went wrong", code="ERROR_CODE")
+        data = json.loads(result[6:-2])
+
+        assert data["error"] == "Something went wrong"
+        assert data["code"] == "ERROR_CODE"
+
+    def test_format_sse_error_without_code(self) -> None:
+        """Test formatting error without code."""
+        result = format_sse_error("Error message")
+        data = json.loads(result[6:-2])
+
+        assert data["error"] == "Error message"
+        assert "code" not in data
+
+
+class TestStreamBuffer:
+    """Tests for StreamBuffer class."""
+
+    async def test_buffer_basic(self) -> None:
+        """Test basic buffer operations."""
+        buffer = StreamBuffer(max_size=10)
+
+        # Producer
+        async def produce():
+            await buffer.put(StreamChunk(id="1", delta="Hello"))
+            await buffer.put(StreamChunk(id="2", delta=" World"))
+            await buffer.done()
+
+        # Consumer
+        chunks = []
+        asyncio.create_task(produce())
+
+        async for chunk in buffer:
+            chunks.append(chunk)
+
+        assert len(chunks) == 2
+        assert chunks[0].delta == "Hello"
+        assert chunks[1].delta == " World"
+
+    async def test_buffer_error(self) -> None:
+        """Test buffer with error."""
+        buffer = StreamBuffer()
+
+        async def produce():
+            await buffer.put(StreamChunk(id="1", delta="Hello"))
+            await buffer.error(ValueError("Test error"))
+
+        asyncio.create_task(produce())
+
+        with pytest.raises(ValueError, match="Test error"):
+            async for _ in buffer:
+                pass
+
+    async def test_buffer_put_after_done(self) -> None:
+        """Test putting after done raises."""
+        buffer = StreamBuffer()
+        await buffer.done()
+
+        with pytest.raises(RuntimeError, match="closed"):
+            await buffer.put(StreamChunk(id="1", delta="Test"))
+
+
+class TestStreamToString:
+    """Tests for stream_to_string function."""
+
+    async def test_stream_to_string_basic(self) -> None:
+        """Test converting stream to string."""
+        async def mock_stream():
+            yield StreamChunk(id="1", delta="Hello")
+            yield StreamChunk(id="2", delta=" ")
+            yield StreamChunk(id="3", delta="World")
+            yield StreamChunk(
+                id="4",
+                delta="",
+                finish_reason="stop",
+                usage=UsageStats(prompt_tokens=5, completion_tokens=3),
+            )
+
+        content, usage = await stream_to_string(mock_stream())
+
+        assert content == "Hello World"
+        assert usage is not None
+        assert usage.prompt_tokens == 5
+
+    async def test_stream_to_string_no_usage(self) -> None:
+        """Test stream without usage stats."""
+        async def mock_stream():
+            yield StreamChunk(id="1", delta="Test")
+
+        content, usage = await stream_to_string(mock_stream())
+
+        assert content == "Test"
+        assert usage is None
--- a/mcp-servers/llm-gateway/uv.lock
+++ b/mcp-servers/llm-gateway/uv.lock