What Google opened
Google has put an early-access Model Context Protocol server in front of Google Home, giving MCP-compatible AI clients a standard way to inspect a user's smart-home environment and act on connected devices. Google's documentation exposes tools to list homes and resources, read current device states, query historical events and execute supported device actions.
Access is rolling out first to Google Home Premium Advanced subscribers in the United States, according to Google launch reporting carried by TechCrunch and The Verge. The Advanced plan costs $20 a month or $200 a year in the US. Setup still has a developer-facing layer: users need a Google Cloud project, the Home API enabled, OAuth credentials and an MCP-capable client. Google documents setup paths for Antigravity, Claude Cowork and OpenClaw.
The useful surface is state, history and control
The MCP server can answer questions grounded in live telemetry, such as whether devices are connected or which lights are on, and can search past state changes and event logs over a time range. It can also run parameterised actions on supported devices, so an agent can move from describing a home to changing it through the same interface.
There are important product gaps. Google's documentation says creating and managing automations through Home MCP is still coming later, and some traits remain experimental. Access to familiar-face data requires a separate consent flow from a manager of the Google Home structure.
Google puts hard limits around physical actions
Google warns explicitly that connecting a real home gives the chosen AI agent control over devices and access to home data. Home MCP applies rate limits and blocks sensitive actions such as unlocking doors, but Google says an agent can still behave unexpectedly or undesirably. Users can revoke the connection through Google Home or their Google account, and Google advises informing other household members when an agent is connected to a shared home.
Those controls set a narrower boundary than the phrase 'control your home' can imply. The server exposes a defined command surface rather than unrestricted access to every capability of every device, and the agent remains subject to the permissions and safety rules Google applies to that surface.
Why the launch matters, and what is still limited
MCP gives third-party agents a common interface to a physical environment that already aggregates Nest hardware, Matter devices and other Works with Google Home products. That makes Google Home a data and action layer that external assistants can use rather than requiring every agent developer to build a separate integration for each device family.
The launch remains early access. Google has not announced a timetable for lower subscription tiers or additional markets, and its own documentation lists latency and experimental-trait issues. The practical test will be whether external agents can use the interface reliably enough for routine household tasks without creating unacceptable privacy, permission or control failures.