MCP Server

When this skill is loaded, announce it to the user by outputting: 🛠️✅ SKILL mcp-server loaded!

Load this skill when working on:

• MCP server setup and configuration
• MCP tool registration and implementation
• API client integrations for MCP tools
• Adding new MCP tools

---

Key Files

File	Purpose
mcp/src/server.py	FastMCP server instance + tool registration
mcp/src/main.py	Entry point
mcp/src/api_client.py	Backend API connection pool (httpx async)
mcp/src/tools/strategy_tools.py	Strategy validation tools

Client Layer

File	Purpose
mcp/src/clients/base.py	BaseAPIClient (shared async HTTP)
mcp/src/clients/data_client.py	Market data queries
mcp/src/clients/backtesting_client.py	Backtest execution
mcp/src/clients/training_client.py	Model training operations
mcp/src/clients/strategies_client.py	Strategy management
mcp/src/clients/operations_client.py	Operation tracking
mcp/src/clients/indicators_client.py	Technical indicator queries
mcp/src/clients/system_client.py	System health/metadata

---

Architecture

Claude Desktop / MCP Client
    │
    ▼ MCP Protocol (stdio)
FastMCP Server (mcp/src/server.py)
    │
    ├── @mcp.tool() decorated functions
    │
    ▼ Domain-specific async clients
Backend API (http://localhost:8000)

Framework: FastMCP (Python MCP SDK)

Safety: Research-only tools. No live trading or order execution.

---

Available Tools

@mcp.tool()
async def hello_ktrdr() -> str:
    """Health check / connectivity test."""

@mcp.tool()
async def check_backend_health() -> dict:
    """Backend connectivity + service status."""

@mcp.tool()
async def get_available_symbols() -> list[str]:
    """List configured trading symbols."""

@mcp.tool()
async def validate_strategy(path: str) -> dict:
    """Validate strategy YAML (v2/v3 format)."""

---

Adding a New Tool

1. Create or find the appropriate client in mcp/src/clients/
2. Add tool function in mcp/src/server.py or a tools file:

   @mcp.tool()
   @trace_mcp_tool()
   async def my_tool(param: str) -> dict:
       """Tool description shown to Claude."""
       client = get_api_client()
       return await client.my_endpoint(param)
   

3. The tool auto-registers via the @mcp.tool() decorator

---

Configuration

Supports multiple backend targets via environment:

• Local dev: http://localhost:8000
• Preprod: Configurable via environment

Client pooling managed by get_api_client() factory using httpx async sessions.