On September 16, Google quietly turned its entire smart home into something Claude and ChatGPT can dial into directly. Home MCP is a Model Context Protocol server that exposes Nest cameras, thermostats, and every Matter-compatible light bulb or plug in your house to outside AI agents — not through a Google-built chatbot, but through whichever assistant you already pay for. TechCrunch reported that early access went live that day, and Unite.AI’s coverage confirmed the same list of supported clients Google named in its own announcement.

That list is the interesting part. Claude, ChatGPT, Google’s own Antigravity, Hermes, and OpenClaw can all now request the same device data and the same control actions, through the same server. That’s not a partnership. It’s an open door — the first time a major consumer IoT platform has handed device control to outside AI agents at this scale, rather than keeping it locked inside a single assistant.

Quick Summary: What Happened

Detail	Info
Date	September 16, 2026 — early access opened
What launched	Home MCP, a Model Context Protocol server for Google Home devices
Supported AI clients	Claude, ChatGPT, Google Antigravity, Hermes, OpenClaw
Access requirement	Google Home Premium Advanced ($20/month or $200/year)
Rollout scope	US only, English only
Setup	Requires a Google Cloud project, OAuth credentials, and explicit MCP permission grants
What agents can do	Review Nest camera history and summaries, monitor device state, execute control actions — not unlock doors

Bottom line: Google just made Nest and Matter devices callable by any MCP-compatible AI agent, gated behind a $20/month tier and a Cloud console setup most consumers have never touched. The capability is real. So is the exposure.

What Actually Happened

The announcement came from Taylor, Google’s Group Product Manager for Google Home & Nest, posted to the Google Home & Nest Community forum rather than a splashy keynote. That’s worth noting on its own — a capability this broad shipped as a developer-forum post, not a stage announcement. Home MCP runs at home.googleapis.com/mcp, and per Google’s developer documentation, it hands a connected agent five things: the ability to list your home’s structure, discover devices and their control schemas, monitor real-time state, execute parameterized control commands, and pull historical event logs.

Translate that out of API language: an AI agent with Home MCP access can tell you what your thermostat is set to, turn your lights on by room name, and — this is the one that’ll get attention — pull up a summary of what your Nest camera saw. Google’s own example, cited in Unite.AI’s writeup, has a parent asking an agent what the kids did when they got home from school, and getting back a summary stitched together from every camera in the house, clips included.

That’s a genuinely different category of capability than “turn off the lights.” It’s a home surveillance archive, queryable in natural language, by a third-party AI model that Google didn’t build.

Which AI Agents Can Control Google Home Devices?

1. Claude — Anthropic’s assistant, connecting via Claude’s MCP support (already covered in our guide to MCP servers).
2. ChatGPT — OpenAI’s assistant, using the same MCP client pattern as its other tool integrations.
3. Google Antigravity — Google’s own agentic coding environment, extended here beyond code into home control.
4. Hermes — listed by Google among early supported clients for Home MCP.
5. OpenClaw — the open-source personal agent platform that already runs with broad permissions across messaging and files; now it can add “control the house” to that list.

None of that access is free-floating. It’s gated behind Google Home Premium Advanced, the $20-a-month tier that replaced Nest Aware earlier this year, per Google’s own announcement of the restructured subscription. Google AI Ultra subscribers get Advanced bundled in automatically; AI Pro subscribers get the cheaper Standard tier and have to pay the difference to unlock MCP access. Everyone else pays the full $20 directly, on top of whatever they’re already paying Anthropic or OpenAI for the agent making the calls.

Setup isn’t a toggle in the Google Home app, either. You need an actual Google Cloud project, the Home API enabled through APIs & Services, an OAuth consent screen configured for external users, and OAuth 2.0 credentials with redirect URIs matching whichever client you’re connecting — Claude, Antigravity, and OpenClaw each need their own. That’s infrastructure work, not a settings menu. It’s the kind of setup that filters out casual users almost by design, and it’s also the kind of setup a lot of people will get wrong.

Why This Matters

Every other major AI-to-smart-home integration up to this point has run through a single company’s own assistant. Amazon’s Alexa controls Amazon-approved devices. Apple’s Siri works inside Apple’s Home app. Google itself has run Assistant-based home control for a decade. What’s new here isn’t AI controlling smart devices — it’s Google letting someone else’s AI do it, at protocol level, without building a Google-branded interface in between.

That’s the MCP thesis playing out at physical scale. We wrote about MCP crossing 97 million monthly SDK downloads back in March, mostly in the context of databases, code repos, and enterprise SaaS. Home MCP is the protocol showing up somewhere with real-world stakes attached — cameras, locks, thermostats, garage doors. When an MCP integration breaks in a coding tool, you get a bad diff. When one misfires in a smart home, something in your physical house does the wrong thing.

Google clearly thought about this. The server enforces rate limits, and it explicitly blocks certain actions — door unlocking chief among them — regardless of what the connected agent asks for. That’s the correct call, and it’s more restraint than some enterprise MCP deployments have shown. But “blocks unlocking doors” is a narrow guardrail, not a permissions system. Garage door openers, smart locks marketed as “smart latches” rather than deadbolts, and any Matter device that doesn’t fall cleanly into “lock” category aren’t obviously covered by that one carved-out exception.

What Are Your Options Now

If you’re a Google Home Premium Advanced subscriber curious about this, the setup is genuinely technical — a Cloud console, OAuth credentials, client IDs — closer to what we described in our MCP servers tutorial than anything in the consumer Google Home app. Don’t attempt it expecting a toggle switch. Expect to spend real time in the Google Cloud console.

If you’re deciding whether to grant an agent camera access, separate the two capabilities in your head before you grant anything. Device control (lights, thermostat) and camera history access are different risk categories, and Home MCP bundles both behind the same permission grant. There’s currently no way to hand an agent “control my thermostat” without also handing it “review my camera archive,” if both are things you’ve connected.

If you share a household with people who didn’t sign up for this, tell them. Google’s own guidance says as much — anyone connecting an agent should inform other people in the home that a third-party AI can now access device data and camera summaries. That’s not a formality. It’s the difference between “I set this up” and “everyone in this house knows an AI model can watch camera footage of them.”

If you’re not on Premium Advanced and have no interest in Cloud console setup, none of this changes anything for you yet. This is early access, US-only, English-only — not a default behavior change to existing Google Home accounts.

The Bigger Picture

The pattern to watch isn’t Google specifically — it’s what happens when IoT platforms start treating “AI agent” as just another client type, the same way they’d treat a phone app or a voice assistant. Once one major platform does this at Google’s scale, the pressure on Amazon and Apple to answer with their own agent-facing MCP equivalents gets real. Neither has announced anything comparable yet. Google clearly wants to be first.

There’s also a quieter signal in how Google separated this. The consumer-facing Home MCP that controls your devices is a distinct product from the “Home Developer MCP” that feeds documentation and Matter specs to coding tools like Antigravity, Claude Code, and GitHub Copilot in VS Code. Google built two different doors — one for AI agents that help you build smart home integrations, one for AI agents that operate your actual house. Keeping those separate was the right instinct. Whether every third-party client respects that separation as cleanly as Google’s own docs do is a different question, and one nobody outside Google can currently audit.

Our Take

We think the capability is legitimately useful and the rollout is legitimately underbaked on security communication. Google did the hard part — building granular device schemas, OAuth-scoped access, rate limiting, and an explicit block on lock control — and then explained almost none of the actual risk tradeoff to the people this affects. A forum post from a product manager is not where you disclose that a third-party AI model, one Google didn’t build and doesn’t control, can now generate a written summary of who walked through your front door and when.

The $20/month gate reads less like a security measure and more like a business decision that happens to double as a filter — it keeps this contained to people already paying for the highest camera-history tier, which narrows the blast radius but doesn’t address it. Rate limits and a no-unlock rule are good engineering. They’re not the same thing as a clear answer to “what happens if the agent I connected gets prompt-injected by something in a video it’s summarizing.” MCP’s own security track record — prompt injection through tool inputs, permission scoping that’s coarser than it should be — hasn’t been fully solved a year and a half after the protocol shipped. Pointing that same class of tool at cameras and door sensors instead of a database raises the stakes without any evidence the underlying protocol issues got fixed first.

Our recommendation: if you’re setting this up, connect device control before you connect camera history, and give it a few weeks before you hand any agent your Nest archive. Early access exists precisely so problems like this surface before the wider rollout. Let someone else find them first.

Frequently Asked Questions

What is Google Home MCP?

Home MCP is a Model Context Protocol server, launched in early access on September 16, 2026, that lets AI agents like Claude, ChatGPT, Google Antigravity, Hermes, and OpenClaw monitor and control Google Home devices — including Nest cameras, thermostats, and Matter-compatible hardware — using natural language instead of a dedicated app.

Which AI agents work with Google Home MCP?

Google named five supported clients at launch: Claude, ChatGPT, Google Antigravity, Hermes, and OpenClaw. Any MCP-compatible agent can technically connect once it’s configured with the right OAuth credentials, but these five are the ones Google explicitly tested and documented.

How much does Google Home MCP cost?

Home MCP access requires a Google Home Premium Advanced subscription, priced at $20 per month or $200 per year. Google AI Ultra subscribers get Advanced included automatically; AI Pro subscribers get the cheaper Standard tier by default and need to upgrade separately.

Can AI agents unlock my doors through Home MCP?

No. Google’s documentation states that Home MCP enforces rate limits and explicitly prohibits sensitive actions, including unlocking doors, regardless of which connected agent requests it.

Is Google Home MCP available outside the US?

Not yet. Early access is limited to English-language Google Home Premium Advanced subscribers in the United States. Google hasn’t announced a timeline for expanding to other countries or languages.

Is Home MCP the same as the developer tools Google offers for Antigravity and Claude Code?

No. Home MCP controls actual devices in your home. A separate product, Home Developer MCP, feeds Matter specifications and API documentation to coding tools like Google Antigravity, Claude Code, Cursor, and GitHub Copilot — that one helps developers build smart home integrations, not operate your house.

Do I need technical skills to set up Home MCP?

Yes, more than the average smart home feature requires. Setup involves creating a Google Cloud project, enabling the Home API, configuring an OAuth consent screen, and generating client credentials for whichever AI agent you’re connecting — closer to setting up a developer integration than flipping a setting in the Google Home app.

---

Last updated: September 20, 2026. Sources: Google Home & Nest Community — Introducing Home MCP · Google Developers — Home MCP · TechCrunch — Your AI agents can now control your Google Home devices · Unite.AI — Google Opens Home MCP Early Access to AI Agents for Smart Home Control · Google — Introducing Google Home Premium · Google Store — Google Home Premium.

Related reading: Model Context Protocol: The USB-C Standard for AI · MCP Hits 97M Installs: The Agent Standard Is Set · Google Antigravity Review 2026 · OpenClaw Review 2026 · Claude MCP Servers Tutorial