mcp-server: Instructions file for GitHub Copilot

Quick context for AI coding agents

This repository implements the Openapi.com MCP gateway (FastAPI + FastMCP). The server acts as a proxy that forwards a client's Bearer token to downstream Openapi services and exposes MCP tools implemented under src/openapi_mcp_sdk/apis/.

Key files

• src/openapi_mcp_sdk/main.py — application entry point. Mounts the MCP app and contains HTTP endpoints /callbacks and /status/{request_id}. Shows how token query params are converted into an Authorization header.
• src/openapi_mcp_sdk/mcp_core.py — central FastMCP instance (mcp), helper make_api_call(ctx, method, url, ...), and processPolling for handling async callbacks.
• src/openapi_mcp_sdk/memory_store.py — in-memory + Memcached-backed callback result store. Exposes get_callback_result and set_callback_result and defines BASE_URL, callbackUrl and memcached config from env vars.
• src/openapi_mcp_sdk/apis/ — each module defines one or more tools decorated with @mcp.tool. Importing these modules triggers tool registration.

What matters for code changes

• Tools are registered when the apis modules are imported by main.py. Don't remove those imports unless you know how to re-register tools.
• make_api_call relies on extracting the Authorization header from FastMCP internals or ctx.request_context.request.headers. Many tools assume the client supplied Authorization: Bearer <token> (or ?token= query param).
• Async APIs use a callback workflow: downstream APIs accept a callback object with url and custom (see company.py). Callbacks are saved with set_callback_result and read with get_callback_result.

Common patterns and code examples

• Registering a tool (example from company.py):
  • Add @mcp.tool above an async function with ctx: Context as last arg.
  • Use make_api_call(ctx, "GET", url) to proxy authenticated requests.
• Creating an async callback request (company full profile):
  • Build custom_context = {"request_id": getSessionHash(ctx), ...}
  • POST to the downstream .../callback endpoint with callback.url = memory store callbackUrl and callback.custom = custom_context.
  • Use processPolling(ctx, request_id, ...) to wait/poll stored results.

Environment and runtime

• Run locally: python src/openapi_mcp_sdk/main.py (README shows python main.py from project root; in this layout use the module path). The server binds to 0.0.0.0:PORT (reads PORT env var, default 80).
• Virtualenv: repository uses uv in README but pyproject.toml shows requires-python = ">=3.13". Note: README states Python 3.9+. If you modify runtime or CI, confirm the correct Python target.

Developer workflows (concrete)

• Start server (local):
  • Create venv and install requirements (README): uv venv && source .venv/bin/activate && uv pip install -r requirements.txt
  • Run: python src/openapi_mcp_sdk/main.py (or use uvicorn as in if __name__ == '__main__').
• Docker: see Dockerfile and README.
• Debugging tips:
  • Check printed logs: main.py and mcp_core.make_api_call print request details and errors.
  • Use /status/{request_id} to inspect async request state set by callbacks.
  • If memcached calls fail, memory fallbacks provide predictable behavior.

Conventions and gotchas for agents

• Do not add or hardcode secrets (tokens/keys) into code — this project explicitly expects the client to supply Bearer tokens. Keep token handling in middleware or test fixtures.
• When adding a new API/tool module:
  • Follow the @mcp.tool decorator pattern and ensure the module is imported from main.py or otherwise registered during startup.
  • Prefer to call make_api_call for HTTP interactions so the common auth/header extraction is reused.
• The code expects ctx to contain request_context.request.headers in some places; use defensive checks when reading headers if you add new code paths.

What I couldn't verify automatically

• Exact developer commands for uv usage and environment variants are described in README; confirm whether you prefer python -m venv flows or uv for reproducibility. Also confirm the Python version (pyproject vs README).