2026-06-16

SharkClean MCP Turns a Robot Vacuum into a Claude-Controlled Device

For Claude and Claude Code users, this is a neat example of where MCP starts to feel genuinely useful rather than just theoretical. Instead of wiring an LLM to some abstract API demo, this project plugs Claude into a real household device: a Shark/SharkNinja robot vacuum that can be started, paused, docked, or sent to specific rooms in plain language.

Key Points

sharkclean-mcp is an MCP server for SharkClean / SharkNinja robot vacuums.
It lets any MCP client, including Claude Code and Claude Desktop, control the vacuum in natural language.
Supported actions include:
- list robots
- check status
- start whole-home cleaning
- clean specific rooms
- pause, resume, stop
- return to dock
- set suction power
- locate the robot by making it beep
The project is unofficial and not affiliated with SharkNinja.
It talks to SharkNinja’s current cloud backend, called “skegox,” rather than the older Ayla cloud.
That matters because newer robots migrated off Ayla; on Ayla they may appear permanently offline even though the app still works.
The README says this server works directly with skegox, which is why it can control newer models that sharkiq-based tools cannot.
Multi-robot accounts are supported, and each tool can take an optional robot name or device ID.
start_cleaning and clean_rooms verify that the robot is online before issuing commands.
Setup requires uv and Python 3.10 or newer.
First-time auth uses a browser-based SharkClean login flow because headless password login is blocked by Auth0.
Tokens are cached locally and refreshed automatically.
Claude Code can be configured with a claude mcp add command, or via .mcp.json.
Claude Desktop can be pointed at the server with a small JSON config.
There’s also a command-line interface, shark-mcp-rc, for scripting and cron jobs.
The repo includes tests, a doctor command for live cloud checks, and notes about limitations and security handling.

My Take

What strikes me is that this is exactly the kind of MCP project that makes the protocol feel practical. A robot vacuum is a small thing, sure, but it’s also a very believable one: “vacuum the kitchen,” “send it home,” “what’s the battery?” are the sort of tasks that map cleanly to natural language and don’t need a human to babysit them.

I think the most interesting part here is not the vacuum control itself, but the backend archaeology. The README is unusually candid about the fact that older community tooling targets Ayla, while newer devices live on SharkNinja’s “skegox” cloud. That is the real story: the project isn’t just an MCP wrapper, it’s a compatibility bridge for devices that had fallen out of reach of older integrations.

I’d be a little cautious about the auth flow, though. The browser-based one-time login is workable, but it’s also a reminder that consumer IoT clouds are fragile, hostile to automation, and not built with developer ergonomics in mind. The README is honest about that, which I appreciate. If this were my setup, I’d probably test it with one robot first, verify the live map and room naming, and then use Claude Code to trigger cleaning in a few scripted routines rather than treating it like a toy voice assistant.

What I like most is the restraint: the server refuses to fire commands if the robot is offline, validates room targets against the live map, and explicitly warns that SharkNinja could change the backend at any time. That’s the right tone for unofficial reverse-engineered infrastructure. It’s useful, but it’s not magic.

The takeaway: this is a strong little example of MCP doing something concrete, useful, and slightly delightful. Not hype — just a well-scoped bridge between Claude and a real device I’d actually want to control from text.

Reference: GitHub - a-funk/sharkclean-mcp: MCP server for SharkClean / SharkNinja robot vacuums — start cleanings, clean rooms, dock, and check status from any MCP client

同じ著者の記事