Google has opened an Early Access bridge that lets personal AI agents inspect and control a Google Home. The Home MCP can list rooms and devices, read live states, query device history and issue commands. That is a meaningful expansion beyond asking Gemini inside Google's own apps, but it is not a consumer-friendly switch to turn on.
The current setup requires Google Home Premium Advanced, a Google Cloud project, Home API access and OAuth credentials for an MCP-compatible client. In the US, Advanced costs $20 per month or $200 per year. Google names Antigravity, Claude Cowork and OpenClaw as supported examples, but the connection process still reads like a developer preview.
It controls devices, but not automations
Google documents five tools for the Home MCP:
| Tool | What an agent can do |
|---|---|
| Home discovery | List the homes a Google account can access |
| Resource discovery | List rooms, devices, traits and available commands |
| Live state | Read current connectivity and device states |
| History | Query past state changes and events over a chosen period |
| Actions | Send supported commands to one or more devices |
Creating or managing Google Home automations is not supported yet. Google also says its safeguards prohibit sensitive actions such as unlocking doors. Those boundaries make the preview narrower than the phrase "AI agent for your home" suggests, but the remaining access is still substantial. An agent can learn what devices exist, whether they are online, what they did and, for supported traits, operate them.
This is a cloud-authorized route into the Google Home graph, not a new Matter feature on an accessory. A Matter light or sensor already added to Google Home may be represented through the account's home resources, but the MCP does not improve that device's Matter clusters, local reliability or compatibility with another ecosystem.
The setup is the first warning label
The official instructions require creating a Cloud project, obtaining access approval, enabling the Home API, configuring an external OAuth consent screen, creating a web application client and publishing the app before connecting an assistant. That friction is useful context: Google is offering a powerful integration surface to experimenters, not replacing the Google Home app for an ordinary household.
The bigger concern is what happens after connection. Google warns that an agent can produce unexpected or unwanted behavior even with rate limits and blocked sensitive actions. It tells users to inform other household members because the agent can control devices and access home data. Access can be revoked later in Google Home or the Google account, and familiar-face data needs a separate consent step.
Do not connect the primary household simply to see what an AI agent can do. A separate test home or a small set of non-critical devices is the more sensible place to evaluate it, followed by a deliberate review of the assistant's own approval controls and data handling. Everyone sharing the home should know what has been authorized.
For most Google Home owners, the right move is to wait. Home MCP becomes more practical when setup no longer requires a Cloud project, Google documents a stable availability path beyond Early Access, and automation management arrives with clear approval boundaries. Today it is interesting infrastructure for technically confident Advanced subscribers, not a reason to upgrade the household's subscription on its own.
Content feedback
Spot a mistake or missing context?
Send a quick note so it can be checked against the source material.