An MCP server for SharkClean / SharkNinja robot vacuums. Lets any MCP client (Claude Code, Claude Desktop, etc.) start cleanings, clean specific rooms, send the robot home, and check status — agentically, in plain language.

Unofficial — not affiliated with or endorsed by SharkNinja.

All commands go through SharkNinja's current cloud ("skegox") — the same backend the SharkClean app uses — so the robot behaves exactly as if driven from the app, with no hardware modification.

Why not the Ayla cloud? The widely-used sharkiq SDK (and the Home Assistant Shark IQ integration built on it) talk to SharkNinja's legacy Ayla IoT cloud. Newer robots (Matrix, AI Ultra, Mach 2, the RV2500+ generation) have been migrated off Ayla onto the skegox backend. On Ayla those robots show up permanently "Offline" with a frozen, stale room map, even though the app controls them fine. This server talks to skegox directly, which is why it works with the newer robots that sharkiq can't reach. The reverse engineering of the skegox API is owed to the shark2mqtt project.

Tools

start_cleaning and clean_rooms first check the robot is online and refuse if it isn't, so commands don't silently vanish. Multi-robot accounts are supported — every tool takes an optional robot argument (name or device id). With one robot, omit it.

Setup

Requires uv and Python ≥ 3.10.

git clone https://github.com/a-funk/sharkclean-mcp
cd sharkclean-mcp
uv sync
cp .env.example .env       # add your SharkClean app email + password
uv run shark-mcp-auth      # one-time browser login (see below)
uv run shark-mcp-doctor    # confirm: prints your robot + current room map

One-time login (shark-mcp-auth)

SharkNinja's Auth0 tenant blocks headless password logins, so the first login happens in your browser using the SharkClean app's own PKCE flow:

1. shark-mcp-auth opens the real SharkClean login page.
2. Log in. The browser then tries to redirect to com.sharkninja.shark://... and usually shows an error, blank page, or an "Open SharkClean?" dialog — that's expected.
3. Copy that full com.sharkninja.shark://... URL (from the address bar, browser history, or the DevTools Network tab with "Preserve log" on) and paste it at the prompt.

That's it. Tokens are cached at ~/.config/shark-mcp/tokens.json (mode 600) and refreshed automatically; you won't need the browser again. See SECURITY.md for how credentials are handled.

Register with your MCP client

Claude Code (user-wide):

claude mcp add --scope user sharkclean -- uv run --directory /ABSOLUTE/PATH/TO/sharkclean-mcp shark-mcp

Or copy .mcp.json.example to .mcp.json, set the absolute path, and Claude Code will pick it up automatically when opened in this directory.

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "sharkclean": {
      "command": "uv",
      "args": ["run", "--directory", "/ABSOLUTE/PATH/TO/sharkclean-mcp", "shark-mcp"]
    }
  }
}

Credentials come from environment variables or the .env file: SHARKCLEAN_EMAIL, SHARKCLEAN_PASSWORD, and optionally SHARKCLEAN_REGION=eu.

Example prompts

• "Start a max-power clean."
• "Vacuum the kitchen and the living room."
• "Is the shark done? What's its battery at?"
• "Send the robot back to its dock."

Command-line control

Beyond the MCP server, shark-mcp-rc drives the robot from a shell (handy for scripts/cron):

uv run shark-mcp-rc status
uv run shark-mcp-rc rooms
uv run shark-mcp-rc clean Kitchen "Living Room"
uv run shark-mcp-rc start max
uv run shark-mcp-rc dock

Development

uv run pytest            # unit tests (fake robot, no cloud calls)
uv run shark-mcp         # run the server on stdio
uv run shark-mcp-doctor  # live end-to-end check against the cloud

See CONTRIBUTING.md. If your robot works, add it to CONFIRMED_MODELS.md.

Notes & limitations

• Auth: the one-time browser login (shark-mcp-auth) yields an Auth0 refresh token, cached at ~/.config/shark-mcp/tokens.json. The server mints short-lived skegox id_tokens from it. SharkNinja's Auth0 tenant rotates refresh tokens, so the server persists the rotated token on every refresh — if you see invalid_grant, re-run shark-mcp-auth.
• Headless/scripted password login is blocked by Auth0 bot detection ("Suspicious request requires verification"); the browser flow is the reliable path.
• Room targeting (clean_rooms) requires a robot with a completed map; older models without room support report no rooms.
• This is unofficial, built on community reverse engineering. SharkNinja can change the backend at any time.