Internal Tools

One hosted MCP server that maps internal API endpoints to agent tools, generated from OpenAPI specs via FastMCP from_openapi(). No per-agent tool wrappers, no duplicated work across repos.

For the MVP it fronts a mock "build-metrics" FastAPI service in this repo; swap API_BASE_URL to point at a real internal service.

Run

uv sync

# terminal 1 — mock backend (port 8000)
uv run uvicorn mock_api.main:app --port 8000

# terminal 2 — MCP server (port 8001)
uv run python mcp_server/server.py

The MCP endpoint is http://localhost:8001/mcp (streamable HTTP).

Connect an agent

Claude Code (this repo already has .mcp.json):

{
  "mcpServers": {
    "internal-tools": { "type": "http", "url": "http://localhost:8001/mcp" }
  }
}

For a team deployment, host mcp_server/server.py on an internal box and point the URL there — every agent gets the same toolset from one place.

Point at a real service

API_BASE_URL=https://your-service.internal uv run python mcp_server/server.py

Requirements on the backend:

• serves its OpenAPI spec at /openapi.json (FastAPI does this out of the box; for non-FastAPI services, hand-write a minimal spec and load it from a file)
• good operation_ids and endpoint docstrings — they become tool names and descriptions, which is what the agent reasons over

Auth: add headers to the httpx client in mcp_server/server.py, e.g. httpx.AsyncClient(base_url=..., headers={"Authorization": f"Bearer {token}"}).

Curating the toolset

route_maps in mcp_server/server.py controls what gets exposed. The MVP excludes /admin/* and anything tagged internal; everything else becomes a tool. Add RouteMap entries (regex, HTTP method, or tag based) to include or exclude endpoints as the toolset grows.

Adding more backends later

The goal is one server, many services: fetch each service's spec, build a FastMCP.from_openapi() server for it, and mcp.mount(other, prefix="svc") them onto the main server so tools stay namespaced.