sgthun/FastMCP

FastMCP (incubator)

FastMCP is a tiny helper to expose selected FastAPI routes as an MCP server using the Streamable HTTP transport. It’s written to be copied out as a standalone library later.

Highlights:

• Introspects a FastAPI app and turns routes into MCP tools
• Filter by tags to build different MCP “personas” on different ports
• Optional approval workflow: queue tool calls for human approval via a pluggable store and a minimal portal

This incubator lives under FastMCP/ in this repo for bootstrapping and will be extracted to its own package.

Quick start (in this repo)

• The entrypoint FastMCP/http_server.py loads the kitchen FastAPI app and exposes an MCP server on port 8091 by default.
• Configure via env:
  • FASTMCP_HTTP_PORT: port to listen on (default 8091)
  • FASTMCP_TAGS: comma-separated FastAPI route tags to include (optional)
  • FASTMCP_REQUIRE_APPROVAL: true/false to enable approval queue (optional)

When approvals are enabled, a simple portal is mounted under /portal to approve/decline/modify queued requests.

Adding per-route MCP metadata

Attach attributes to your route endpoint functions to customize the exposed tool:

• mcp_enable: bool (default True)
• mcp_name: str (defaults to route name or path-derived)
• mcp_description: str
• mcp_tags: Iterable[str] (merged with FastAPI route tags)
• mcp_requires_approval: bool (overrides global default for this route)

Example:

@router.post("/do_something", tags=["automation"])
async def do_something(payload: Payload):
    do_something.mcp_name = "do_something_tool"
    do_something.mcp_description = "Does something important"
    do_something.mcp_tags = ["automation", "safe"]
    do_something.mcp_requires_approval = True
    return {"ok": True}

Approval store

• InMemoryApprovalStore provides a simple testing store.
• Implement the ApprovalStore interface to back the queue with your datastore of choice.

Notes

• This MCP server proxies calls back into the FastAPI app using its TestClient. For production, consider delegating to the app’s HTTP API over the network.
• The tool input schema is a generic object. Future versions may infer request models to generate JSONSchemas.
• This is an incubator; APIs may change when extracted.