Fastmcp Caddy Proxy

A two-container runtime: a FastMCP application server, and a reverse proxy sitting in front of it that enforces bearer-token auth before traffic reaches the MCP server.

Layout

mcp-app/         FastMCP server (Streamable HTTP transport, port 8000, internal-only)
proxy/           Caddy reverse proxy (bearer-token auth, port 8080, exposed)
docker-compose.yml

• mcp-app — a single-file server.py using PEP 723 inline script metadata, run with uv run server.py. It's not published to the host network; only proxy can reach it, over the mcp-net compose network.
• proxy — Caddy checks the Authorization: Bearer <token> header against MCP_PROXY_TOKEN and rejects anything else with 401 before reverse-proxying to mcp-app:8000.

Running it

cp .env.example .env
# edit .env and set MCP_PROXY_TOKEN to a real value, e.g.:
#   openssl rand -hex 32
docker compose up --build

The MCP endpoint is then available at http://localhost:8080/mcp, and requires:

Authorization: Bearer <MCP_PROXY_TOKEN>

Any request without a matching header gets a 401 from the proxy without ever reaching the FastMCP app.

Local dev without Docker

cd mcp-app
uv run server.py

This runs the FastMCP server directly (no proxy/auth) on http://localhost:8000/mcp.

Notes

• The two example tools (ping, echo) are placeholders — replace them with real tools in mcp-app/server.py.
• MCP_PROXY_TOKEN is a shared secret for this local/dev setup. For anything beyond local dev, manage it with fnox and swap the Caddy static-token check for something stronger (OAuth, per-client tokens, etc.) before exposing it outside a trusted network.
• Lint with ruff check . from the repo root (config in pyproject.toml).