Roborock MCP Server

Remote MCP server that lets Claude (claude.ai, Claude Desktop, Claude mobile) control a Roborock vacuum via natural language. Runs as a Python serverless function on Vercel.

Architecture

Claude (claude.ai / Desktop / mobile)
        │ HTTPS JSON-RPC
        ▼
api/mcp.py  (FastAPI, single Vercel serverless function)
        ├── /authorize, /token, /.well-known/oauth-*   OAuth shim
        │     claude.ai's connector UI requires OAuth; these endpoints wrap
        │     our one static MCP_AUTH_TOKEN in OAuth shape. The "Client ID"
        │     you paste into Claude's connector settings must equal
        │     MCP_AUTH_TOKEN, so this adds no less security than a plain
        │     bearer check.
        └── /api/mcp   JSON-RPC endpoint
              ├── initialize   – protocol handshake
              ├── tools/list   – returns tool schemas
              └── tools/call   – dispatches to roborock_client.py
                       │
                       ▼
              roborock_client.py
                       │ loads a cached login session (no live login —
                       │ this account requires 2FA email-code, which can't
                       │ run inside a serverless function)
                       ▼
              python-roborock device_manager → MQTT → Roborock cloud → vacuum

Related MCP server: congatudo_mcp

Repo layout

roborock-mcp/
├── api/
│   └── mcp.py           FastAPI app: JSON-RPC + OAuth shim, single entrypoint
├── roborock_client.py   Wraps python-roborock calls used by the MCP tools
├── test_auth.py         One-off script: log in (email code), list devices, cache session
├── requirements.txt     Runtime deps (kept in sync with pyproject.toml)
├── pyproject.toml       Vercel's Python build reads dependencies from here
├── vercel.json          Routes /authorize, /token, /.well-known/* to api/mcp.py
├── .env.local           Local-only secrets (gitignored, not committed)
└── roborock_session.json  Cached login session (gitignored, not committed)

Tools exposed

Local setup

python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

Create .env.local:

ROBOROCK_EMAIL=you@example.com

(Password login isn't supported for accounts with 2FA enabled — leave ROBOROCK_PASSWORD unset and use the email-code flow.)

Phase 1 — prove auth works

python test_auth.py

Logs in (prompts for the emailed code on first run), lists your real devices, and caches the session to roborock_session.json so subsequent runs skip the login prompt.

Phase 2/3 — run the MCP server locally

python -m uvicorn api.mcp:app --reload

Test with curl or the MCP Inspector against http://127.0.0.1:8000/api/mcp.

Windows note: both test_auth.py and roborock_client.py set asyncio.WindowsSelectorEventLoopPolicy() — the default Proactor loop doesn't support add_reader/add_writer, which the MQTT session needs.

Phase 4 — deploy to Vercel

vercel login
vercel env add ROBOROCK_EMAIL production
vercel env add ROBOROCK_USER_DATA production   # paste roborock_session.json contents
vercel env add MCP_AUTH_TOKEN production        # your own random secret
vercel --prod

ROBOROCK_USER_DATA exists because Vercel's filesystem isn't writable/persistent between invocations — the cached session has to come from an env var instead of the local roborock_session.json file.

Phase 5 — connect to Claude

Settings → Connectors → Add custom connector

• URL: https://roborock-mcp.vercel.app/api/mcp

• OAuth Client ID: your MCP_AUTH_TOKEN value

• Client Secret: leave blank

Then enable the connector for a conversation and try "what's my vacuum's status".

How it works (plain-English walkthrough)

You type "start cleaning" to Claude → Claude has no way to touch your vacuum directly, it only knows how to call "tools" some server exposes. This project is that server, sitting between Claude and Roborock's cloud.

1. Logging in. Roborock's cloud needs proof you own the account before it'll let anyone control your vacuum. Logging in is just a web request: send email + password (or email + one-time code), get back a session token — like a wristband at a concert, show it once and you're in for a while without re-checking ID. We cache that token (roborock_session.json) so we're not forced to log in fresh on every single request (especially since this account requires an emailed code, not just a password).

2. Finding your devices. Another web request, logged in this time, says "list my home's devices" and gets back your vacuum's name, model, and unique ID.

3. Talking to the vacuum is not a normal web request. Normal websites are request-then-response. Vacuums use MQTT instead — think of it as a radio channel. Your vacuum is tuned to a channel; our server tunes into the same one. To send a command, we broadcast a message on that channel and the vacuum (listening) acts on it. The vacuum also constantly broadcasts its own status ("battery 87%", "currently cleaning") on that same channel, which is how we read status back — not an instant reply, but "tune in, ask, wait for the broadcast."

Why not plain HTTP? HTTP can't have the vacuum speak up unprompted — you'd have to keep re-asking "any updates?" MQTT lets the device push updates on its own.

4. python-roborock (the library). Someone already wrote all the login/MQTT/decoding logic as a reusable package — we didn't reinvent it, just call its functions: client.pass_login(...) logs in, create_device_manager(...) finds devices and opens the MQTT connection, device.v1_properties.command.send(RoborockCommand.APP_START) sends "start cleaning" over the radio, status.refresh() asks for status and waits for the broadcast reply.

5. Our server. roborock_client.py is a thin wrapper — it names the specific actions Claude is allowed to trigger (start, pause, dock, status, consumables, rooms, clean a room) and calls into python-roborock for each. api/mcp.py is the actual web server: when Claude wants to use a tool, it sends a request in JSON-RPC format (a standard shape for "call this function with these arguments, give me the result"), e.g.:

{"jsonrpc": "2.0", "id": 5, "method": "tools/call",
 "params": {"name": "start_clean", "arguments": {}}}

Our server reads method/params.name, runs the matching function, and replies with the same id so Claude can match the reply to its request:

{"jsonrpc": "2.0", "id": 5,
 "result": {"content": [{"type": "text", "text": "{'result': 'cleaning started'}"}]}}

6. Putting it on the internet (Vercel). Your own computer isn't always on and has no public address Claude can reach. Vercel hosts the code for you at a permanent URL (https://roborock-mcp.vercel.app), spinning it up briefly per request rather than running an always-on machine you'd have to manage.

7. The login problem on Vercel. Vercel's environment doesn't keep files between requests, so the cached "wristband" file doesn't survive there. Its contents get copied into an environment variable (ROBOROCK_USER_DATA) instead — a permanent setting Vercel keeps around for the app to read, rather than a file on disk.

8. Securing the server. The URL is public, so anyone who finds it could send commands to your vacuum unless we gate it. MCP_AUTH_TOKEN is a random secret every request must include, or it's rejected as unauthorized.

9. The OAuth complication. Claude's own connector setup screen insists on an OAuth flow (the "Sign in with Google"-style dance) and won't accept a plain secret directly. Real OAuth: app redirects you to the provider's login page, you type your password there (the app never sees it), provider redirects back with a temporary code, app exchanges that code for an access token behind the scenes. We built a minimal fake version of just the code-exchange steps — no real login screen, since there's only one user (you) and no separate identity to protect. If the "Client ID" you type into Claude's connector settings matches MCP_AUTH_TOKEN, we auto-approve and hand back that same token as the "access token." This satisfies Claude's UI requirement without weakening security below a plain bearer check.

Known limitations

• No auto re-login. If the cached Roborock session expires, get_status etc. will start failing with an auth error. Fix: rerun test_auth.py locally (email-code login), then update the ROBOROCK_USER_DATA env var on Vercel and redeploy.

• clean_room is untested — get_rooms returned an empty list on this account (map may not be fully synced yet). The command path is implemented per the confirmed RoborockCommand.APP_SEGMENT_CLEAN API, but hasn't been exercised against a real room id.