MCP Tool Server

A production-oriented MCP (Model Context Protocol) server that exposes tools to AI agents, built on FastMCP + FastAPI, with Docker support.

Built as a learning/portfolio project referencing the architecture of PrefectHQ/fastmcp, extended with several original pieces:

• Auto-discovery tool registry — drop a file with a register(mcp) function into app/tools/ and it's picked up automatically. No manual wiring in main.py.

• Per-client rate limiting — in-memory token bucket, keyed by API key.

• API key auth on the HTTP surface (/tools/* endpoints), independent from the raw MCP protocol endpoint.

• Structured logging with per-request IDs and timing.

• /health and /metrics endpoints for basic observability.

Tools included

Related MCP server: mcp_server

Running locally

pip install -r requirements.txt
python -m app.main

Server starts on http://localhost:8000.

• MCP protocol clients connect to http://localhost:8000/mcp

• HTTP clients can call POST /tools/{tool_name}/invoke with header x-api-key: <key>

Running with Docker

docker compose up --build

Adding a new tool

Create app/tools/my_tool.py:

def register(mcp) -> str:
    @mcp.tool()
    def my_tool(x: int) -> int:
        """Doubles a number."""
        return x * 2
    return "my_tool"

That's it — no changes to main.py needed, it's discovered on startup.

Tests

pytest tests/

Environment variables