Google Home now has an MCP server, so Claude and ChatGPT can run your house

Until this week, giving a third-party program control of a Google Home meant going through Google: the Assistant, the Works with Google Home program, or the developer APIs behind them. On September 16 Google opened another door. It is an MCP server for Google Home, in early access, and any agent that can call Model Context Protocol tools can connect to it, Claude and ChatGPT included.

Google's announcement, posted to its Home and Nest community and signed by Taylor, Group Product Manager for Google Home and Nest, says "any agent that supports calling MCP tools, such as Google Antigravity, Claude, Hermes, OpenClaw or others, can now interact with all of the devices and event history in the Google Home ecosystem." Sarah Perez's report for TechCrunch adds ChatGPT to the list. It covers Nest doorbells and thermostats and any device connected through Works with Google Home or Matter. The gate is money and geography: early access is rolling out in English to US subscribers of Google Home Premium Advanced, the $20 a month tier.

What an agent gets, and how you hand it over

Google's examples from early testers: ask what the kids did after school and get a summary across every camera with the clips; count the loads of laundry in a week from device state history; have the agent announce over a Home speaker when an asynchronous task finishes; build a custom dashboard in everyday language. The setup is not a toggle. A user creates a Google Cloud project, configures it for Home MCP, hands the configuration details to the agent of their choice, and the agent asks the user to sign in and grant permissions, with guides on the Google Home Developer Center. Google is also shipping Home Developer MCP, a separate server that grounds coding tools in its Home API references, integration guides, and Matter, OpenThread, and Thread documentation, Android Police reports.

Rate limits, no door unlocking, and a warning

The announcement says Home MCP "enforces rate limits and safety protections, such as prohibiting sensitive actions like unlocking doors," and then adds a line most launch posts leave out: "depending on your agent, connecting it to Home MCP can result in unexpected or even undesired behavior." What it does not describe is per-device or per-room scoping of the grant, or an audit log of what an agent did. The permission, as written, is for the ecosystem. Google asks for feedback in its Smart Home for Developers community and gives no date for availability beyond the US or beyond the paid tier.

Why a build studio cares

MCP is how our agents talk to tools, and this is the first server from a platform we would plausibly connect to that has motors and cameras behind its tool list. The engineering question is the one we already ask about payments and email: what is the blast radius of one bad tool call? Google has answered the worst case itself, by refusing to unlock doors at the server, and left the rest to the agent. Our rule for client agents is that any tool that changes the world sits behind an allowlist of specific operations, never a blanket grant, and the Home MCP grant as written is one grant for the whole account. We would connect it read-only first, camera summaries and event history, and put every device write behind a confirmation step. The product work starts when a client asks for the opposite, and the answer has to be a list of exactly which calls the agent may make without asking.

Next step: read Google's announcement, the Home MCP setup guide on the Developer Center, TechCrunch's report, and Android Police's write-up of the developer server. If you are wiring an agent to anything that moves, unlocks, or records, write to us at hello@gattyworks.com.