Google began early access to Home MCP on September 16, opening the smart home to third-party AI agents. Any agent that speaks the Model Context Protocol (MCP) can read the state and event history of devices connected to Google Home, and act on them. The rollout starts with Google Home Premium Advanced subscribers in the US, in English.
Not one assistant, but every agent that speaks MCP
Until now, controlling a smart home meant going through whichever assistant the platform provided. For Google Home, that meant Google's own assistant.
Home MCP changes that. Any agent that supports MCP tool calls can reach the devices and event history inside the Google Home ecosystem. Google cites its own Google Antigravity, Anthropic's Claude, and agents such as Hermes and OpenClaw as examples. In effect, Google is giving up the exclusive position of its own assistant and moving the entry point to an open standard.
The device coverage is broad. Google Nest cameras and doorbells, Nest Thermostats, Matter light bulbs, and generally anything registered with Google Home are included.
Reading the history changes what an agent can do
Google's examples make clear that this is more than a replacement for remote control.
The first is cross-camera analysis. Ask your agent what the kids did after they got home from school, and it will summarize activity across every camera in the house and surface the relevant clips. Chasing footage camera by camera collapses into a single question.
The second is history analysis. By walking through device state history, an agent can count how many loads of laundry ran in a week, or how long the lights were left on. What makes this interesting is that sensor values become a log of daily life rather than raw readings.
The third is voice notification. When an agent is working through a long-running or proactive task, it can deliver an audio message over a Google Home speaker once it finishes. You do not need to be watching a screen to follow progress.
The fourth is dashboard generation. You can have an agent assemble a view tailored to your own home using everyday language.
Google describes all of this as giving an agent real-world physical context. Once the agent you are talking to knows what is happening inside your house, the range of things worth asking it genuinely shifts.
The front door stays locked
Handing permissions to an outside party raises obvious questions about safety.
Home MCP enforces rate limits, and sensitive actions are prohibited. The one Google names explicitly is unlocking doors. Even if an agent misbehaves, the front door will not open.
At the same time, Google warns that connecting an agent can lead to unexpected or even undesired behavior depending on which agent it is, and advises users to review developer policies and terms of service carefully. That caveat is the nature of a system that accepts agents outside Google's control, and the candor is welcome.
A separate MCP server for developers
Alongside the user-facing Home MCP, Google has also introduced Home Developer MCP.
This one grounds coding tools in verified technical documentation: Home API references, integration guides, and material on Matter, OpenThread, and Thread. Supported tools include Google Antigravity, Claude Code, Cursor, and GitHub Copilot in VS Code.
Generated code that assumes outdated specifications or APIs that do not exist is a problem in smart home development too. Pinning the reference material to official documentation is a sensible response.
A 20 USD tier, and setup that takes some work
Early access is limited to subscribers of Advanced, the upper tier of Google Home Premium. Advanced costs 20 USD per month (about 3,100 yen), or 200 USD per year (about 31,000 yen), and includes 60 days of event history and 24/7 continuous recording. Google AI Pro subscribers can move up to Advanced for an extra 10 USD per month (about 1,600 yen), and Google AI Ultra subscribers get it included.
※1 USD = 156 JPY
Setup is not a single toggle in an app. You need to create a Google Cloud project and configure it to use Home MCP. After that, you pass the MCP configuration details to your agent, sign in, and grant permissions. Full instructions are published at the Google Home Developer Center.
For now availability is limited to US English, so there is no way to try it from Japan yet. Still, the broader move of a smart home platform opening itself to agents through an open standard looks likely to spread to other platforms.
Summary
Google has started early access to Home MCP, letting MCP-capable AI agents work with Google Home devices and event history. Intended uses include cross-camera summaries, analysis of household activity logs, voice notifications through speakers, and custom dashboards. Rate limits and a ban on door unlocking cover the safety side, while Google asks users to check the behavior of outside agents themselves. Access is for Google Home Premium Advanced subscribers at 20 USD per month, starting in the US in English.
