Connecting a real home to an AI agent allows that agent to control devices on your behalf, Google warns in developer documentation. “Depending on your agent, connecting it to the Home MCP may result in unpredictable or even unwanted behavior.” This warning can be found in the setup guide for the Google Home MCP.
See also: OpenClaw and Claude store AI-generated podcasts on Spotify
Google opened it to early access on September 16. It allows external AI agents to read and control a Google Home.
Any agent that supports the Model Context Protocol can connect. Google lists Antigravity, Claude, Hermes , and OpenClaw. A connected agent has access to each device and the history of events behind it, including Nest doorbells, thermostats, and Matter light bulbs. Google has documented four things its early testers created.
An agent can summarize footage from cameras. A parent can ask what the kids were doing when they got home and receive the clips back. They can also read the status history of appliances, covering how many loads of laundry were done in a week or how long the lights were left on.
Additionally, an agent completing a lengthy task can send an audio message through a Google Home and create a custom dashboard, replacing the Google Home app interface with one created by the agent. “This gives your favorite AI agent a real physical context,” Taylor Lehman, product manager for Google Home and Nest, wrote in the announcement.
The scope is consumer-grade, but the setup isn’t straightforward. A user creates a Google Cloud project, enables the Home API, and then configures an OAuth consent screen. After that, they need a client ID and secret, the correct redirect URI for their agent, and publish the app. Google suggests asking the agent itself to do the configuration, though the prompt it provides still points to a pre-production sandbox, indicating how early this is. Early access covers Google Home Premium Advanced subscribers in the United States, in English.
See also: Gemini for Google Home: Security cameras enable automations
That’s the $20-a-month tier. TechCrunch asked about other tiers and purchases, but Google didn’t provide details. Automations aren’t currently supported through the Home MCP, and Google lists latency and experimental device features as known issues. Some limits are strict. The Home MCP enforces rate limits and blocks sensitive actions.
Google cites unlocking doors as an example of what an agent can't do, and access can be revoked from the Google Home app at any time. Facial recognition requires a second, separate consent.
A user must be an administrator of the home structure, operate a compatible Nest camera or doorbell with known face detection enabled, and visit a consent link associated with their own client ID and home. Google also advises users to notify other household members. If a home is shared, others need to be aware that the agent can control devices and read home data.
One suggested alternative is to create a separate home for testing. Handing over the keys to a home to an agent presents a known problem. Prompt injection can hide instructions within ordinary content, which the agent then follows. Researchers demonstrated this in early versions of Gemini in the smart home, and Google has since patched these vulnerabilities.
See also: Google Home Gemini 3.1: New complex command capabilities
The security issue shifts from Google to whichever agent the user connects to, and Google acknowledges this in its warning.
