Google Opens Its Smart Home to Any AI Agent With Home MCP

Google opened early access to Home MCP on September 16, 2026, a server built on the Model Context Protocol that lets third-party AI agents work with devices in a user's Google Home ecosystem. The company framed the release as a step toward what it calls agentic smart homes, and it is the first time Google has let rival AI assistants reach the smart home platform it had kept as a Gemini-only experience.

Related reading: Google Gives Its CC Agent a Family Account and a Permission Model · Meta's Muse Is a Personal AI Agent That Wants Your Passwords and Your Purchases.

The integration lets an agent discover a household's devices, read their current states, search device and camera event history, and run supported actions. Google listed its own Antigravity agent, Anthropic's Claude and Claude Cowork, ChatGPT, Hermes and OpenClaw as examples, and said any client that can call MCP tools can connect. The announcement came from Taylor Lehman, group product manager for Google Home and Nest, in a post to the Google Home and Nest community.

What the agents can actually do

MCP, or the Model Context Protocol, is an open standard that Anthropic introduced to connect AI applications to external data sources and tools. Google's implementation exposes five tools, according to its Home MCP developer guide: list_homes to find the households an account can reach, list_home_resources to enumerate devices and other addressable resources, list_home_states to read current states, list_home_history to retrieve past state information, and run_home_actions to execute supported commands.

MCP has moved quickly from a developer convention to a widely adopted interface. Anthropic introduced the protocol in late 2024, and through 2026 it spread from coding tools into enterprise software and, now, physical devices. That trajectory explains Google's decision to support it. A common protocol means the company does not have to build a separate integration for every assistant, and developers do not have to learn a vendor's proprietary function-calling format. The tradeoff is that the same openness makes it easier for users to treat whichever agent they prefer as the interface and the smart home platform as plumbing, which is the position Google is trying to own rather than lose.

In practice, that turns a natural-language request into a sequence of tool calls. A question about whether the lights were left on requires the agent to discover resources and read their state. A request to turn a supported light off adds an action call. Because the tools expose device history, an agent can reason across time, which is where the more interesting uses appear.

TechCrunch described uses such as reviewing camera summaries, monitoring activity and building custom dashboards. The Verge noted that an agent could analyze footage across several cameras or count how many times a particular thing happened at home. Google said an agent can also send an audio message through a Google Home speaker when it finishes a task, which turns the speaker into an output channel for an assistant that is not Google's own.

The limits Google built in

Google has constrained what agents can reach. The company blocks certain sensitive actions, including unlocking doors, and it applies rate limits. It does not yet let agents create or manage automations, which it said will come in a later release. Access to familiar-face data requires separate consent from the manager of the structure, and the structure must contain at least one compatible Nest camera or doorbell with familiar face detection enabled.

Some traits are marked experimental and may not work as expected, and Google warns of longer-than-expected latency while it works on performance. The company also cautions that the behavior of an agent is only as predictable as the agent itself. "Depending on your agent, connecting it to Home MCP can result in unexpected or even undesired behavior," Lehman wrote, recommending that users review Google's developer policies and terms of service before they connect anything.

Those caveats matter because a tool call against a smart home has consequences that cannot be undone by editing a file. A misread state is harmless; an unintended command to a thermostat, appliance or lock is not. Google's decision to withhold unlock actions is a recognition that the permissions model needs to be narrower than the capability of the underlying devices.

What it costs and how to set it up

Home MCP is not a consumer toggle. Access at launch is limited to Google Home Premium Advanced subscribers in the United States, in English, and Google said the rollout will continue over the coming weeks. The Advanced tier costs $20 per month or $200 per year and already includes 60 days of event-based video history, descriptive notifications and daily summaries. There is no separate charge for the MCP integration.

Setup is aimed at developers and technically confident users. It requires an existing Google Home environment, a Google Cloud project with the Home API enabled, an OAuth consent screen configured for an external audience, and an OAuth client with the home.platform.v2 scope. The server URL is https://home.googleapis.com/mcp. Google published step-by-step guides for Antigravity, Claude Cowork and OpenClaw. The requirement to create a Cloud project and manage OAuth credentials puts the feature well beyond a simple settings screen, at least during early access.

Google also launched a second, lighter server. The Home Developer MCP is free and needs no user OAuth flow. It grounds coding tools such as Antigravity, Claude Code, Cursor and GitHub Copilot in VS Code in Google's Home API reference, the Matter 1.5.1 specification and the Thread 1.4.1 specification, so developers build against documentation rather than whatever a model remembers from training.

Google's infrastructure play

The strategic reading of the announcement is less about smart home convenience than about position. Google retired Google Assistant and replaced it with Gemini for Home, which remains the default assistant in the Home app and on Nest speakers. Home MCP does not replace it. It adds a second layer through which outside agents can interact with the same devices and data.

That is a deliberate choice to become the infrastructure that other companies build on. Google opened API access to its smart home in 2024, and its developer messaging has increasingly framed Google Home as a platform rather than only a product. The Verge compared the ambition to the role AWS plays for internet businesses. If the devices and event history are the substrate, then Google benefits when developers connect any agent at all, because every connection still runs through Google Home.

The move also lands in a crowded field of personal AI assistants. Meta launched Muse, its personal agent, earlier in September, and Snap introduced an anticipatory assistant alongside its new Specs glasses. Opening the smart home to outside agents gives Google a way to stay central as users choose their preferred assistant rather than assuming they will use Gemini.

Privacy, safety and the trust question

Giving an AI agent access to cameras, locks, thermostats and appliances raises questions that a software integration normally does not. Google's answer is a combination of scoped tools, rate limits, blocked actions and separate consent for recognizable-face data. The company also warns that behavior depends on the agent, which effectively shifts part of the safety burden to whichever third-party assistant a user chooses.

There is a platform-history reason for caution as well. Google has launched and retired a long list of smart home developer programs, including Android Things, Works with Nest and Project Brillo, and developers have learned to weigh long-term commitment before building on the next one. The company will have to convince them that Home MCP is infrastructure rather than an experiment.

What to watch

Three questions will shape how this develops. The first is whether Google expands Home MCP beyond the United States and beyond premium subscribers, which determines whether it becomes a mass-market feature or stays a developer preview. The second is whether the company adds automations and whether it keeps door unlocking permanently off limits. The third is whether other smart home platforms respond with their own MCP servers, which would turn the protocol into the common interface for controlling a home through an agent.

For now, Google Home is the first major smart home platform to hand third-party AI agents a defined set of tools for reading and acting on a home. The early access is narrow, the setup is technical and the safety model is still being written, but the direction is clear. The home is becoming a set of discoverable resources that an agent can query and, within limits, command.

Cover: Wikimedia Commons, "Google Home with Home Hub and Home Mini on table" by Y2kcrazyjoker4, licensed CC BY-SA 4.0, cropped to 16:9.