Google Home MCP server enters early access, letting Claude, ChatGPT, Hermes, OpenClaw and Antigravity control home devices

On September 16, 2026, Google announced early access to Home MCP — a Model Context Protocol (MCP) server for the Google Home ecosystem. Announced by Taylor Lehman, Group Product Manager for Google Home & Nest, in a Google Home & Nest Community post titled "Introducing Home MCP: enabling your agent to interact with your home," the server lets AI agents that support calling MCP tools — Google named Google Antigravity, Claude, Hermes, OpenClaw or others — "interact with all of the devices and event history in the Google Home ecosystem," from Google Nest doorbells and thermostats to Works with Google Home (Matter) light bulbs. TechCrunch additionally reported ChatGPT among the launch-capable agents.

The early-access rollout started the same day for Google Home Premium Advanced subscribers in the US, in English, and continues over the following weeks. Google also shipped a companion Home Developer MCP server (first documented in July 2026) that grounds coding tools such as Antigravity, Claude Code, Cursor and GitHub Copilot in the Home API reference, the Matter 1.5.1 spec and the Thread 1.4.1 spec.

Illustrative use cases Google published at launch: cross-camera analysis ("what did the kids do when they got home from school?" with a summary across all cameras plus clips), home-history analysis (how many loads of laundry in a week, how long lights were left on), voice interaction (the agent sends an audio message over a Google Home speaker when an async task finishes), and building custom smart-home dashboards.

• Before: Controlling a Google Home smart home through AI meant Google's own stack — the Home app and Gemini for Home (Google Assistant's successor) — or third-party hacks such as Home Assistant's MCP integration. Rival AI assistants had no sanctioned, first-party channel into Nest cameras, Nest thermostats, or Matter devices.
• Change: Google exposed the Google Home Graph (structures, devices, states, event history, actions) as a first-party, OAuth-scoped MCP server. Any agent that can call MCP tools can now, with owner consent, query and control the home — including Anthropic's Claude, Nous Research's Hermes, OpenClaw, Google's own Antigravity, and (per TechCrunch) OpenAI's ChatGPT.
• After: The physical home becomes an agent-controllable surface independent of assistant vendor. Google positions itself as the neutral "infrastructure layer" of the agentic smart home; the consumer choice of assistant is decoupled from the smart-home platform. Agent access to the home is monetized via the top subscription tier (Home Premium Advanced, $20/month or $200/year), and safety rails (rate limits, no door unlocks, separate familiar-face consent, revocable access) are enforced at the server boundary.

Home MCP is an MCP server acting as a proxy between smart-home infrastructure and AI applications. Per the official user guide (developers.home.google.com/mcp/home):

• Server endpoint: https://home.googleapis.com/mcp (SSE transport), OAuth scope https://www.googleapis.com/auth/home.platform.v2. (The doc's sample configuration prompt also mentions a preprod sandbox URL https://preprod-home.sandbox.googleapis.com/mcp; the manual configuration for Antigravity/Claude/OpenClaw uses the production home.googleapis.com/mcp.)
• Five executable tools:
  1. list_homes — structure discovery: homes/structures accessible to the authenticated user.
  2. list_home_resources — device/area/trait/attribute/command-schema enumeration.
  3. list_home_states — real-time connectivity and trait state.
  4. list_home_history — historical state changes and event logs over time ranges (camera/doorbell activity included).
  5. run_home_actions — execute parameterized commands on target devices, subject to Google Home restrictions.
• Setup flow (documented, developer-grade): create a Google Cloud project → enable the Home API → configure an External-audience OAuth consent screen → create a web-application OAuth client ID with per-client redirect URIs (Antigravity: https://antigravity.google/oauth-callback; Claude Cowork: https://claude.ai/api/mcp/auth_callback; OpenClaw: user-specified local redirect URI) → publish the app → paste client ID/secret into the agent of choice, then the agent drives OAuth login and permission grant. Google publishes step-by-step guides for Antigravity, Claude Cowork, and OpenClaw, and offers a copy-paste "automated setup" prompt for agents that manage MCP configs from chat.
• Access control & safety: rate limits; sensitive actions such as unlocking doors are prohibited on the server; familiar-face data requires a separate, explicit consent link (structure manager + at least one compatible Nest camera/doorbell with familiar-face detection enabled); agent access can be revoked from the Google Home app or My Accounts page. Google's own warning: "depending on your agent, connecting it to Home MCP can result in unexpected or even undesired behavior."
• Role separation: Gemini for Home remains the default assistant in the Home app and on Nest speakers; Home MCP is an additional, opt-in layer. The companion Home Developer MCP (https://homedevelopers.googleapis.com/mcp) is a separate read-only server exposing one search_documents tool over indexed corpora (Google Home Developer Center, OpenThread, Matter 1.5.1, Thread 1.4.1), with an unauthenticated shared-quota trial path or an API-key authenticated path (Home Developers API). Its search_documents tool is annotated read-only and requires citing returned source_url values.

1. The physical home is now an agent surface. For the first time, a major first-party smart-home platform gives rival AI assistants sanctioned access to real devices, camera history and event logs. As The Verge put it, the change is not "Claude can turn lights on/off" — it is that an agent of choice has access to the underlying data and control layer of the home, a step toward a "whole home computer" that understands, reasons across history, and acts.
2. MCP wins the physical-world connector layer. Google — which has every incentive to keep users inside Gemini — wired its home ecosystem into a protocol it does not own (created by Anthropic in November 2024 and now supported by OpenAI, Microsoft, xAI and others). This is the strongest signal yet that MCP is the de-facto standard for agent-to-tool connectivity, extending from code editors to the physical world. The first major first-party server where a tool call has consequences beyond software.
3. Google bets on the "AWS of the smart home" B2B2C strategy. After the 2024 Home API launch and the Gemini for Home full-stack offering, Home MCP lets third parties build consumer experiences on Google's infrastructure — even if those experiences use Claude, Hermes or OpenClaw rather than Gemini. It converts platform loyalty from "assistant brand" to "infrastructure brand."
4. Monetization of agentic access. Agent access is gated behind the top subscription tier ($20/mo or $200/yr), continuing Google Home Premium's conversion of Nest Aware into a broader subscription and attaching an agent-control premium.
5. Sets competitive pressure. Google now leads Apple Home and Amazon Alexa on third-party-agent openness (neither supports MCP clients per The Robotics Media); it also legitimizes the Home Assistant/OpenClaw-style MCP approaches that hobbyists had already built.

• Ask your agent of choice: "What did the kids do after school?" and get a cross-camera summary with clips, without opening the Google Home app.
• "How many loads of laundry did we do last week?" / "How long were the lights left on?" — answered from device-state history.
• "Turn off all the outside lights," "Is my home secured?" — real-time state queries and parameterized commands via natural language.
• Agents deliver async results as voice messages through Google Home speakers.
• Custom smart-home dashboards built by the agent from home structure and event data.
• Coding agents (Antigravity, Claude Code, Cursor, GitHub Copilot) answer Home API/Matter/Thread questions grounded in verified docs instead of stale training data (Home Developer MCP, live-verified in this run).
• Developers can, for the first time, build "bring your own assistant" smart-home products on Google's graph — an architectural precedent for agent-IoT.

• When does Home MCP reach non-US markets, and will Standard-tier subscribers ever get access?
• When do automations (create/manage) become MCP tools — and are they gated by extra consent?
• Is ChatGPT officially supported? Google named only Antigravity, Claude, Hermes, OpenClaw; TechCrunch/Thurrott add ChatGPT. (MCP's open design means any standard client can connect; the question is first-party support status.)
• Which action categories beyond door unlocks are server-blocked? (Garage, alarms, emergency calls reported by media, not confirmed by Google.)
• What does the rate-limit envelope actually look like for production agent workloads?
• Will Google add outbound/speaker voice delivery as a first-class tool with structured APIs, or is it agent-mediated only?
• Does Home MCP expand to Home APIs' 600M+ device base (e.g., cloud-to-cloud integrations) or stay scoped to consumer Google Home structures?
• How will the week's agent-safety incidents shape Google's own guardrail roadmap for Home MCP?

• Short term (weeks): staged rollout to all eligible US Premium Advanced subscribers; feedback cycles via the Smart Home for Developers community; latency fixes; more experimental traits stabilized.
• Medium term (months): automations support ("coming soon") — the biggest functional unlock; possible expansion beyond US English; possible ChatGPT documentation/formal support; likely rate-limit/productization tuning.
• Longer term (2027): watch for (a) Alexa/Apple MCP responses, (b) service-provider/proptech adoption of agentic Home access, (c) the first high-profile security incident shaping server-side safety norms, and (d) whether Google's B2B2C "infrastructure layer" bet converts the developer community despite the platform graveyard history.
• Narrative arc: Home MCP fits the week's meta-story — agents moving from code to workflows to the physical world, with safety rails (rate limits, blocked actions, consent gates) as the differentiator against the same week's agent-unsafety headlines.

The headline — "Claude and ChatGPT can now run your Google Home" — undersells the story. Home MCP is the moment the physical home became a standard agent surface, and the moment Google chose infrastructure over assistant lock-in: it wired its smart-home graph into a protocol created by its rival Anthropic and handed consumer choice of assistant to the user. It is simultaneously a counterpoint to the week's agent-risk panic: here is a major lab-grade deployment demonstrating what "safe-by-construction" physical agent access can look like — scoped OAuth, rate limits, a hard no on door unlocks, separate consent for face data, and explicit warnings that agent behavior may still surprise. The honest caveats (US-only, $20/mo gate, developer-grade setup, no automations yet, ChatGPT naming ambiguity) keep this a promising early access rather than a finished revolution. For the AI-news audience: this is the story that makes the agent hype physical — in both the empowering and the unsettling sense. Watch the rollout, watch the automation tools, and watch whether this finally becomes the Google smart-home platform that sticks.