mcp-agent-hub: Skill for Claude Code

Verifying MCP Agent Hub changes

The hub is a FastAPI + FastMCP server (mcp_hub/hub.py) with a browser dashboard at /, REST under /api/*, and the MCP endpoint at /mcp. hub.db is CWD-relative (D27), so verify against a scratch instance launched from a throwaway directory — never against the repo root unless you intend to touch the real DB.

Launch a scratch instance (port 8001, isolated DB)

# from any scratch dir (its hub.db lands there):
& <repo>\venv\Scripts\python.exe -m uvicorn --app-dir <repo> mcp_hub.hub:app --host 127.0.0.1 --port 8001

Ready when GET http://127.0.0.1:8001/api/state returns 200. The venv is venv\ at repo root (not .venv). The real deployment uses python run_hub.py (supervisor, port 8000) — check nothing is already on :8000 before assuming the field is clear.

Drive the real surfaces

• MCP tools (through the middleware, real streamable-HTTP):

  from fastmcp import Client
  async with Client('http://127.0.0.1:8001/mcp') as c:
      await c.call_tool('register_agent', {'agent_id': 'probe', 'skills': []})
  

• SSE stream: curl -sN --max-time 30 http://127.0.0.1:8001/api/events → expect one data: snapshot immediately, a new one after any mutation, : keepalive comments when idle (20 s).
• REST mutations: POST /api/broadcast ({"payload": "..."}) is the easiest state-change trigger; hostile-origin probes (-H "Origin: http://evil.com" or a bad Host:) must 403 on /mcp and /api/*.
• Dashboard: open http://localhost:8001/ via claude-in-chrome; confirm panels update live (Refresh = "Live") with only two /api/ requests in the network log (/api/state first paint + the /api/events stream). Mutate via MCP/REST and screenshot — panels must change without a reload.

Gotchas

• httpx ASGITransport (used by tests/test_mcp.py) buffers whole bodies — it can never consume /api/events; the SSE tests use a real uvicorn on an ephemeral port (live_server fixture) for this reason. Don't "simplify" them back.
• Module-global asyncio primitives (hub.notifier) bind to the first event loop that uses them; tests get a fresh one per test via the autouse fixture.
• PowerShell backgrounds long pytest/uvicorn invocations; pipe through Select-Object -Last N shows nothing until the process exits — a silent run may be a hung SSE consumer, not a slow suite.