Home MCP is the name of the door Google just opened. On 16 September 2026 the company began rolling out early access to a Model Context Protocol server for the Google Home ecosystem, which lets any AI agent that can call MCP tools read your device states, search your event history and switch things on and off on your behalf.
That is a genuine change in who holds the controls. Until now, the assistant inside your house was Google’s assistant. Gemini for Home answered from the Home app and from Nest speakers, and everything else talked to your lights through a narrow set of partner integrations. With this server live, a third-party agent — Anthropic’s Claude, OpenAI’s ChatGPT, Hermes, OpenClaw or Google’s own Antigravity — can sit in the same seat.
The rollout is far narrower than the headline suggests. Access is limited to Google Home Premium Advanced subscribers in the United States, a tier that costs $20 a month, and the setup involves creating a Google Cloud project before your agent sees a single light bulb. There are rate limits, and there is a list of actions the server simply refuses.
This article walks through what the server exposes, who can reach it today, how the connection is made, what Google has deliberately walled off, and what the whole move tells you about where artificial intelligence assistants are heading inside physical spaces.
Table of contents
- What Home MCP Actually Is
- The Seven Essential Facts About Home MCP
- Who Can Use Home MCP Right Now
- Which AI Agents Can Call Home MCP
- How to Connect an Agent to Home MCP
- What Home MCP Will Not Let an Agent Do
- Home MCP Versus the Home Developer MCP Server
- The Security Questions Home MCP Raises
- Why Google Shipped Home MCP Now
- What Home MCP Means If You Build or Buy Smart Home Technology
- Frequently Asked Questions About Home MCP
- References and Further Reading
What Home MCP Actually Is
Strip away the branding and this is a protocol adapter. Model Context Protocol is the open standard Anthropic published in 2024 for connecting AI models to external tools and data, and it has since been adopted across the industry as the default way an agent reaches a system it does not own.
A server, not a new assistant
Home MCP does not replace Gemini for Home. Google’s own assistant remains the primary interface across the Home app and Nest speakers, and nothing about that changes. What the new server adds is an additional control layer that other software can address directly, using the same tool-calling grammar an agent already uses for a calendar or a code repository.
The four capability groups
Google’s developer documentation describes the server around four things an agent can do: enumerate the rooms and devices in a structure, query live device connectivity and status, execute parameterised actions on those devices, and read past states and event logs. Google’s phrasing for the whole package is blunt — “connect agents natively to physical spaces.”
Why event history is the interesting half
Most coverage of Home MCP focused on switching lights, which is the least interesting thing on that list. The history access is the part with novelty. An agent that can read your camera event log can answer questions no rule-based automation ever could: how long a light was on, how many laundry loads finished this week, what happened across three cameras in one sequence.
What “any agent” really means
Google’s claim is that any agent supporting MCP tool calls can interact with every device and every event record in the ecosystem — from a Nest doorbell and thermostat to any “Works with Google Home” or Matter-compatible bulb. The compatibility promise is therefore about the protocol, not about a partner list. If your agent speaks MCP, Home MCP will answer it.
The Seven Essential Facts About Home MCP
Before the detail, here is the announcement compressed into the facts that decide whether this affects you.
| # | Fact | Detail as announced |
|---|---|---|
| 1 | Launch date | 16 September 2026, rolling out over the following weeks |
| 2 | Status | Early access, not general availability |
| 3 | Who qualifies | Google Home Premium Advanced subscribers, US, English |
| 4 | Price of that tier | $20 per month or $200 per year |
| 5 | Named agents | Antigravity, Claude, Hermes, OpenClaw, ChatGPT |
| 6 | Setup requirement | A Google Cloud project with the Home APIs enabled |
| 7 | Hard limit | Sensitive actions blocked, including unlocking doors |
Every one of those rows narrows the story. A headline that says Google will let any agent run your smart home is accurate about the protocol and generous about everything else.
The gap between the protocol and the rollout
There is no announced timeline for other subscription tiers, other countries or other languages. Nothing in Google’s materials commits to moving Home MCP out of early access at all. Treat the current release as a controlled experiment with paying volunteers rather than a shipped platform feature.
Who Can Use Home MCP Right Now
The subscription gate is the single biggest constraint, so it is worth understanding what Google Home Premium actually sells before deciding whether Home MCP is within reach.
The two paid tiers
Gemini for Home works without paying — basic voice control and some Ask Home features are free. The paid tiers add camera intelligence. Standard is $10 a month or $100 a year and brings Gemini Live, descriptive alerts and 30 days of event-based video history. Advanced is $20 a month or $200 a year and adds daily home summaries plus 60 days of event history and 10 days of 24/7 recording.
| Feature | Free | Premium Standard | Premium Advanced |
|---|---|---|---|
| Monthly price | $0 | $10 | $20 |
| Annual price | $0 | $100 | $200 |
| Event video history | Limited | 30 days | 60 days |
| 24/7 video history | None | None | 10 days |
| Descriptive alerts | No | Yes | Yes |
| Daily home summaries | No | No | Yes |
| Home MCP early access | No | No | Yes (US) |
What the annual prices are worth
Both tiers price the annual plan below twelve monthly payments. Standard at $10 a month runs to $120 a year against a $100 annual plan, a saving of $20. Advanced at $20 a month runs to $240 against $200, a saving of $40. Both work out at the same 16.7% discount, which is the arithmetic a buyer should check before assuming the annual plan is the better deal for a feature still in early access.
The geography and language gate
Early access is United States only and English only. There is no stated plan for the United Kingdom or the European Union, where camera-derived event history carries heavier data-protection obligations than it does in the US. A UK household reading about Home MCP today is reading about something it cannot switch on.
Which AI Agents Can Call Home MCP
Google named five agents in its materials, and the list is more revealing than it first appears.
Google named its competitors
Antigravity is Google’s own agent suite. Claude is Anthropic’s. ChatGPT is OpenAI’s. Hermes and OpenClaw are independent. A platform holder listing two direct rivals as supported clients on day one is a deliberate signal that Home MCP is being positioned as neutral infrastructure rather than a Gemini feature.
Nothing here is an integration deal
Because the server speaks a public protocol, none of those agents needed Google’s permission and none of them needed to sign anything. Any client that can load an MCP server configuration is a candidate. That is the practical meaning of “any agent” — an open door with a subscription check on it, not a partner programme.
What an agent is likely to do well
The early use cases Google and reporters described cluster around synthesis rather than switching: summarise camera footage across rooms, work out how long the lights ran, count completed laundry cycles, push a voice message through a Home speaker, or assemble a custom dashboard from live device state. Those are natural language processing tasks over structured event records rather than control problems, which is exactly where a language model is strong.
How to Connect an Agent to Home MCP
The onboarding is developer-shaped, and that is the honest reason this is early access rather than a toggle in the Home app.
The prerequisites
You need an active Google Home structure with devices on it, a Premium Advanced subscription, access to a Google Cloud project, and an MCP-capable client. Three of those four are unremarkable. The Cloud project is the one that will stop most households.
The configuration steps
Reports of the setup guide describe creating the Cloud project, enabling the Home APIs on it, configuring the OAuth consent screen, creating a web-application OAuth client ID with the right redirect URIs, and publishing the app. Only then do you hand the MCP configuration details to your agent and ask it to connect, signing in and granting permissions on the way through.
| Step | Where it happens | Who normally does this |
|---|---|---|
| Create a Cloud project | Google Cloud console | A developer |
| Enable the Home APIs | Google Cloud console | A developer |
| Configure OAuth consent | Google Cloud console | A developer |
| Create an OAuth client ID | Google Cloud console | A developer |
| Add the server to the agent | Your MCP client | Anyone comfortable editing config |
| Sign in and grant scope | Google account flow | The homeowner |
Why the friction is probably intentional
An OAuth client you build yourself is a per-household credential. It keeps the blast radius of a mistake inside one home and one Cloud project, and it gives Google a revocation point that does not depend on the agent vendor. Friction here is a safety property, not an oversight — but it does mean Home MCP is currently a hobbyist and developer feature wearing consumer clothes.
What Home MCP Will Not Let an Agent Do
Google put limits on the server before anyone asked, and the limits define the product as much as the capabilities do.
Rate limits and refused actions
Both the announcement and the user guide state that the server enforces rate limits and safety protections, including a prohibition on sensitive actions such as unlocking doors. That single example is well chosen. A smart lock is the one device where a hallucinated tool call has an immediate physical consequence that no software rollback can undo.
Google’s own warning
Google still cautions that connecting outside agents could result in unexpected behaviour. That is an unusual sentence to publish alongside a launch, and it should be read literally rather than as legal boilerplate. The company is shipping a capability it cannot fully predict the use of, and saying so.
The honest limitation list
| Limitation | Practical effect |
|---|---|
| Door unlocking blocked | Locks stay outside agent control |
| Rate limits applied | No high-frequency polling or automation loops |
| Advanced tier only | Excludes free and Standard households |
| US and English only | No UK or EU access at launch |
| Cloud project needed | Not a one-tap consumer setup |
| Early access label | Behaviour and availability can change |
Home MCP Versus the Home Developer MCP Server
Two Google MCP servers with similar names shipped into the same news cycle, and conflating them will waste your afternoon.
One serves your house, one serves your IDE
Home MCP connects a personal agent to a live structure: real devices, real state, real history, authenticated through a user OAuth flow. The Home Developer MCP server at homedevelopers.googleapis.com does something entirely different — it exposes a search_documents tool that returns matching Google Home documentation in markdown, so a coding assistant can ground itself in verified technical docs.
Different authentication, different clients
The developer server authenticates through unauthenticated trial paths or project API keys rather than the user OAuth flow Home MCP requires, and it is aimed at coding tools: the Antigravity suite, Claude Code, Cursor and GitHub Copilot in VS Code.
| Dimension | Home MCP | Home Developer MCP |
|---|---|---|
| What it reaches | Devices and event history | Documentation text |
| Authentication | User OAuth | API key or trial path |
| Typical client | Claude, ChatGPT, OpenClaw | Claude Code, Cursor, Copilot |
| Subscription required | Premium Advanced | None |
| Physical side effects | Yes, it can act | No, read-only text |
Why both exist at once
Together they describe Google’s ambition plainly: be the layer other people’s agents call, whether those agents are running someone’s house or writing the code for the next device that joins it.
The Security Questions Home MCP Raises
Handing a language model the ability to act on physical hardware moves familiar risks into an unfamiliar place, and the MCP ecosystem’s own record is not spotless.
The protocol’s track record
MCP has had a rough security adolescence. Commentators noted flaws across the wider ecosystem, including servers shipped without enforced authentication, and there have been documented breaches since adoption accelerated. None of that is a Google defect, but it is the neighbourhood Home MCP has moved into.
Prompt injection now has a light switch
The specific worry with a house is indirect prompt injection. An agent that reads camera descriptions, notification text and device names is reading attacker-influenceable content, and it is holding tools that change the physical world. The door-unlock prohibition caps the worst outcome; it does not remove the category.
The questions worth asking before connecting
Who else has access to the agent session, what the agent’s provider retains from tool outputs, whether camera event summaries leaving the Google estate is acceptable to your household, and how you revoke the OAuth client if an agent misbehaves. The last one has a clean answer: delete the credential in your own Cloud project.
For anyone deploying agents at work
The same discipline applies to business deployments of autonomous AI agents. Scope the credential narrowly, log every tool call, and keep a human approval step in front of anything with a physical or financial consequence. Home MCP is a consumer product demonstrating an enterprise lesson.
Why Google Shipped Home MCP Now
The timing says as much as the feature does.
The assistant war moved into the house
Every major AI vendor now ships an agent that wants context about your life. Google’s advantage is that it already owns the sensors: doorbells, thermostats, speakers and cameras that produce a continuous stream of real-world events. Exposing that through Home MCP turns a defensive moat into a distribution channel.
Better to be the plug than the appliance
If households are going to pick an agent, Google would rather be the thing every agent connects to than lose the room to whichever assistant wins. Home MCP is the infrastructure play, and it fits a broader push to make Google the platform layer for smart home developers.
The subscription does real work here
Restricting Home MCP to a $20 tier is not only a monetisation choice. It bounds the early user population to people already paying for camera intelligence, which keeps the support load, the abuse surface and the privacy exposure smaller while the behaviour of third-party agents is still an open question.
What Home MCP Means If You Build or Buy Smart Home Technology
Whether or not you can switch it on today, the release changes some planning assumptions.
For device makers
Matter and “Works with Google Home” certification just became more valuable, because certification is what puts a device inside the set an agent can see. A product that only speaks to its own app is now invisible to an entire emerging control layer.
For integrators and IT teams
Expect client questions about connecting agents to building systems well before the tooling is ready for it. The safe answer today is the one Google’s own limits imply: read and summarise freely, act on low-consequence devices, keep locks and access control on a separate, human-approved path.
For households
The practical advice is narrow. If you are in the US on Advanced and comfortable with a Cloud console, Home MCP will show you something genuinely new. If you are not, nothing is lost by waiting, because early access features change and this one has no published roadmap.
For anyone writing an agent
The pattern Google used is worth copying. A public protocol, a user-owned credential, published rate limits and an explicit list of refused actions is a reasonable template for exposing any system with physical consequences to an autonomous caller.
Frequently Asked Questions About Home MCP
Does Home MCP replace Gemini for Home?
No. Gemini for Home remains Google’s primary assistant in the Home app and on Nest speakers. Home MCP adds a parallel control path for external agents.
Can my agent unlock my front door?
No. Sensitive actions are blocked, and door unlocking is the example Google names explicitly.
Do I need to pay to use Home MCP?
Yes. Early access requires Google Home Premium Advanced at $20 a month or $200 a year, in the United States.
Which agents work with it?
Google named Antigravity, Claude, ChatGPT, Hermes and OpenClaw, but any MCP-capable client is a candidate because the server speaks a public protocol.
Is Home MCP available in the UK?
Not at launch. Early access is United States and English only, with no announced timeline for other markets.
How hard is the setup?
Harder than a consumer feature should be. It requires a Google Cloud project, enabled Home APIs and a self-created OAuth client before the agent connects.
References and Further Reading
Your AI agents can now control your Google Home devices
Google Home is going agentic via integration with the MCP standard
Google Home Developers: Home MCP
MCP Reference: homedevelopers.googleapis.com
Google Opens Home MCP Early Access to AI Agents for Smart Home Control
Google opens Home ecosystem to third party AI agents through MCP
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.