Google opened early access to a Model Context Protocol server for Google Home on September 16, letting any MCP-capable AI agent inspect a home's structure, read live device state, issue commands and query event history. The capability list is broad. The exclusions are the more interesting part: the server enforces rate limits and refuses sensitive actions such as unlocking doors.
Key takeaways
- Home MCP exposes five capability areas β structure discovery, resource discovery, state monitoring, device control and historical event analysis β to agents including Claude, ChatGPT, OpenClaw, Hermes and Google Antigravity.
- Access is limited to U.S. subscribers of Google Home Premium Advanced at $20 a month, and Google declined to commit to wider tiers or markets.
- The server blocks sensitive actions such as unlocking doors and applies rate limits, and Google warns that connecting an agent can produce unexpected behavior.
What the server actually exposes
The documented surface is a set of tools an agent calls rather than a chat integration. An agent can enumerate the homes it can reach, list devices along with area layouts, traits, attributes and command schemas through a list_home_resources call, check real-time connectivity and trait values, execute parameterized commands against a target device, and pull past state changes over a chosen window.
That last one matters more than it sounds. Historical event access is what turns an agent from a voice remote into something that can answer questions a dashboard cannot β why the thermostat cycled at 3 a.m., which door opened before the camera clip. TechCrunch reported that the feature set covers reviewing camera summaries, monitoring activity and building custom dashboards, and that the server works with any device in the ecosystem β Nest doorbells and thermostats, plus Works with Google Home and Matter devices such as light bulbs.
Where Google drew the line
The safety posture is the release's clearest editorial statement. Rather than trusting each agent's own guardrails, Google put the restriction server-side: the protocol will not perform the sensitive actions, regardless of which model is asking. Door unlocking is the named example.
Google's own documentation is unusually blunt about residual risk, warning that connecting a real home to an agent can produce unexpected or undesired behavior depending on the agent, and advising users to tell other household members that an agent can now control devices and read home data. Developers are pointed toward creating a separate test home instead. Access is revocable at any time from the Google Home app or the account page.
The setup is a developer workflow, not a toggle
Nobody is enabling this from a phone in thirty seconds. Connecting requires creating a Google Cloud project, enabling the Home API, configuring an OAuth consent screen for an external audience, generating a web-application client ID, and registering the redirect URI that matches the target client β a distinct callback for Antigravity, for Claude, or for a local OpenClaw install. Only then does the agent prompt for sign-in and permissions.
The early-access framing is visible in the endpoint itself, which currently points at a preprod sandbox host under the home.platform.v2 scope. Google is gathering feedback through its Smart Home for Developers Community during the period.
Outlook
Google already runs MCP endpoints across Cloud, its data platforms, developer tooling and Workspace, so the protocol itself is not new territory for the company. What is new is the target: a consumer's physical space rather than a work account. That raises the cost of the failure modes the protocol's own roadmap has been wrestling with, because a mis-fired tool call here turns the lights off in an occupied house rather than corrupting a spreadsheet. Whether the door-lock carve-out stays permanent is the signal worth watching.
FAQ
Which AI agents can connect to Google Home MCP?
Any client that supports the Model Context Protocol. Google names Claude, ChatGPT, OpenClaw, Hermes and its own Antigravity environment, and the documentation lists per-client OAuth redirect URIs for several of them. The connection is authenticated with OAuth credentials the user generates in their own Google Cloud project.
Can an AI agent unlock my door through Home MCP?
No. Google states that the server prohibits sensitive actions such as unlocking doors and applies rate limits, enforcing the restriction at the protocol level rather than relying on the connected agent. Other device categories, including lights, thermostats and cameras, are controllable.
What does it cost and where is it available?
Home MCP early access is rolling out to U.S. subscribers of Google Home Premium Advanced, the $20-per-month tier. Google would not say whether or when access will extend to other subscription tiers or other markets.
