Google Home Just Became an MCP Server

On September 16, Google opened early access to Home MCP, a Model Context Protocol server for the Google Home ecosystem. Any MCP-capable agent — Claude, ChatGPT, Hermes, OpenClaw, Google's own Antigravity — can now monitor devices, review event history, summarize camera footage, and execute control actions across Nest doorbells, thermostats, Matter-enabled gear, and "Works with Google Home" devices.

Sensitive actions stay blocked. Unlocking a door is not on the menu.

Access is US-only, limited to subscribers of the $20-per-month Google Home Premium Advanced tier. Setup requires creating a Google Cloud project, configuring it for Home MCP, handing the resulting configuration to your agent, and granting permissions when prompted.

The strategic read

Google just handed the keys to its smart home to its competitors.

That framing is too cynical, but it is not wrong about the mechanics. Home MCP means a user can run their house through Claude or ChatGPT without touching the Google Home app or talking to Gemini. Google is providing the device layer and letting someone else own the interface.

This is a deliberate position, and it is the same one Google took with Android. When you cannot win the assistant war outright, you win by being the substrate everyone builds on. Google Home has an enormous installed base of Nest hardware and the broadest Matter compatibility story in the industry. Anthropic and OpenAI have the agents people actually want to talk to. Rather than fight that, Google is monetizing the connection at $20 a month.

There is a defensive read too. Apple has just shipped Gemini-powered Siri. Amazon has Alexa+. If the assistant layer is going to commoditize, the company that owns the devices and the protocol endpoint is in a better position than the company that owns the wake word.

Why MCP and not an API

Google Home has had an API for years. Home MCP is not a new capability so much as a new calling convention, and the convention is the point.

MCP standardizes how an agent discovers what a tool can do, inspects its schema, and calls it. An agent that speaks MCP can use Home MCP without anyone writing a Google Home integration for that specific agent. That is the difference between a partnership and a protocol: the first requires a deal per client, the second requires none.

The adoption curve behind this is real. Monid, an open-source tool-routing layer, reported agent transactions doubling from roughly 4 million to more than 8 million between August 31 and September 15. Anthropic has been pushing physical-world MCP. Meta shipped a WhatsApp Business MCP the same week Google shipped this one. The protocol has crossed from an Anthropic convention into industry plumbing, and Google adopting it for consumer hardware is confirmation rather than news.

What it can actually do

The interesting capability is not control. It is event history and camera summarization.

Turning on a light with natural language has been solved, badly, for a decade. What has not been available to a general-purpose agent is context — what happened at this house, when, and in what order. An agent that can query camera events and device activity can answer questions the Google Home app cannot: whether the back door was opened while you were out, whether the thermostat is cycling more than it did last month, whether the dog walker actually came.

That is a materially different product from a voice remote. It turns the house into a data source an agent can reason over, and it is the reason the capability sits behind a paid tier rather than shipping free to everyone with a Nest thermostat.

Custom dashboards are also supported, which is a quiet acknowledgment that the Google Home app's interface has never satisfied power users.

The setup tax is the real gate

Create a Google Cloud project. Configure it for Home MCP. Export the configuration. Hand it to your agent. Grant permissions.

That is not a consumer flow. It is a developer flow shipped to consumers, and it will keep Home MCP in the hands of people who already know what a service account is. Combined with the $20-a-month Premium Advanced requirement and US-only availability, the practical audience for this at launch is small and technical.

That is probably intentional. Handing agents the ability to act on household devices has a failure surface, and a setup process that filters for technical users is a crude but effective form of staged rollout. The blocked sensitive actions — no door unlocking — say the same thing.

The risk nobody has priced

Emergence AI published research the same week testing eight ten-agent environments against prompt injection, misinformation, and memory exposure. Some agents acted on adversarial material up to 46 hours later.

Now consider what Home MCP exposes: camera event history and device activity, read by a general-purpose agent that also reads your email, your calendar, and arbitrary web pages. The injection surface is not the smart home. It is everything else the agent touches, with the smart home as the payload.

Google blocked the obvious catastrophic action. It did not — and structurally cannot — block an agent from being talked into summarizing your camera history into a document that leaves your account. The sensitive-action blocklist protects against the agent doing something physical. It does not protect against the agent leaking something informational.

What to watch

Whether the setup flow gets simplified. A one-tap OAuth handoff would take this from a developer feature to a consumer one, and would multiply both the usefulness and the exposure.

Whether Apple responds. HomeKit has no equivalent, and Apple has just conceded the assistant layer to Gemini. A HomeKit MCP server would be a bigger signal than this one.

The first injection incident involving a physical device. It will happen, it will be reported badly, and it will determine how fast the sensitive-action blocklist shrinks.