Google Hands Your Home to the Agents: Home MCP Opens Smart-Home Control to Claude, ChatGPT and Friends

On Wednesday, September 16, 2026, Google quietly crossed a line that the smart-home industry has been circling for a decade. The company opened early access to Home MCP, a Model Context Protocol server that lets third-party AI agents — Claude, ChatGPT, Hermes, OpenClaw, or Google’s own Antigravity — securely read and act on the devices, sensors and event history of a real, physical home. Your agent of choice can now ask how many lights are in the house, whether the doors are locked, what the kids did when they got home from school, and then turn the thermostat down — all through its own interface, in natural language.

The announcement, signed by Taylor Lehman, Group Product Manager for Google Home & Nest, positions Home MCP as “the next step in unlocking agentic smart homes.” It is also one of the most literal expressions yet of the agentic-AI thesis: a frontier model with authorized, structured access to the control layer of your household.

What Home MCP actually does

Model Context Protocol, the open standard originally pushed by Anthropic, gives AI applications a standardized way to call external tools and read external data. Google had already shipped MCP servers for Google Cloud, its data platforms, developer tools and Workspace. Home MCP extends that pattern to the living room — and notably, to competitors’ agents, not just Gemini.

Google’s MCP servers overview groups Home MCP’s functions into four capability areas:

1. Entity enumeration — list the rooms and devices in a structure (“how many lights are in the house?”)
2. Real-time state monitoring — query live device state such as connectivity (“is the house locked down for the night?”)
3. Device control — execute parameterized actions (“dim the living-room lights to 40%”)
4. Historical analysis — query past states and chronological event logs (“what happened while I was out?”)

The scope covers any device in the Google Home ecosystem: Nest doorbells and thermostats, plus the long tail of “Works with Google Home” and Matter devices like smart bulbs.

The use cases Google highlighted from early testers are pointedly cross-device and temporal — the kind of questions no single-device app can answer. One example: ask your agent what the kids did when they got home, and it returns a summary stitched from every camera in the house along with the relevant clips. Another: use device state history to count how many loads of laundry the household ran last week, or how long the lights were left on. Agents can also send audio messages through Google Home speakers when a background task finishes, and assemble fully custom smart-home dashboards from a natural-language description.

The fine print: $20 a month and a Cloud project

Early access is not casual. It requires a Google Home Premium Advanced subscription — the $20/month or $200/year tier that carries 60 days of event-based video history, 10 days of 24/7 camera history, descriptive notifications and video-history search — rolling out in English to US subscribers over the coming weeks. (Nest Aware has been folded into this branding: legacy Nest Aware Plus subscribers land on Advanced.)

Setup is developer-adjacent rather than consumer-simple. Users must create a Google Cloud project, enable the Home API, configure an OAuth consent screen for an external audience, create a web-application OAuth client ID with redirect URIs for their chosen AI client, and publish the app. The server endpoint is https://home.googleapis.com/mcp with the OAuth scope home.platform.v2, and Google provides step-by-step guides for Antigravity, Claude Cowork and OpenClaw. Familiar-face data from Nest cameras carries an additional explicit consent gate: a home manager must grant access, and only if a compatible camera or doorbell has familiar-face detection enabled.

Google wouldn’t comment on if or when access widens to other tiers or markets, and it is soliciting feedback through its Smart Home for Developers Community during the early-access window.

Alongside the consumer-facing server, Google also launched Home Developer MCP, a companion server that grounds AI coding tools — Antigravity’s CLI and IDE, Claude Code, Cursor, GitHub Copilot in VS Code — in verified documentation: the complete Home API references, the Matter specification, and OpenThread and Thread documentation. That pairing is the strategic tell: personal agents operate the home through Home MCP while developer agents build against it through Home Developer MCP.

Guardrails — and honest warnings

Handing an AI agent the keys to smart locks, thermostats, HVAC and appliances raises obvious safety questions, and Google’s answer is a mix of hard limits and blunt disclosure. Home MCP enforces rate limits and safety protections — most notably, it will not allow an agent to unlock doors. Google describes the system as safe and secure by design, built so agents “cannot take sensitive actions on a user’s behalf.”

But the company is unusually candid about the residual risk. “Connecting it to Home MCP can result in unexpected or even undesired behavior,” Lehman writes, directing users to Google’s developer policies and terms of service. The user guide advises anyone sharing a home with others to inform them that an agent can control devices and access home data, and suggests setting up a separate test home for experimentation. Agent access can be revoked at any time from the Google Home app or the user’s My Accounts page.

Known limitations at launch: some device traits are marked experimental, latency can run longer than expected, and — notably — creating or managing automations through Home MCP is not yet supported, though Google says it plans to add that in a future release.

The real story: Google as the home’s infrastructure layer

The deepest change isn’t that Claude can flip your lights. It’s that a third-party agent now has authorized access to the data and control layer of the home — the substrate for what The Verge’s Jennifer Pattison Tuohy calls a “whole home computer” that can understand what’s happening, reason across history, and act on your behalf. The smart home stops being a command-and-control panel and starts being contextual and proactive.

It’s also a calculated platform play. Google has spent two years positioning itself as the smart home’s infrastructure: opening API access in 2024, making Gemini for Home a full-stack offering, and now exposing the whole thing over an open protocol. As Tuohy observes, the strategy is B2B2C — Google supplies the AI and home infrastructure while other companies build consumer-facing agents and products on top, much as AWS became the substrate for the last generation of software businesses. Even if developers bolt Claude or OpenClaw onto it rather than Gemini, Google still owns the layer underneath.

There is an irony worth naming: this is the same week that AI-safety headlines dominated the news cycle. Google is responding to that environment with a permissioned, rate-limited, consent-gated architecture rather than an open API — arguably the most responsible version of this idea anyone could ship. But the trust question runs deeper than protocol design. Google has killed or abandoned Android @ Home, Weave, Project Brillo, Works with Nest and Google Assistant over the years; developers and users alike will have to decide whether MCP is the platform Google finally sticks with — and whether the agents connecting to it deserve the access they’re being granted.

For now, the smart home has a front door, and it speaks MCP.

---

Cover image generated with MiniMax image generation. Sources are listed in the article metadata.