Google Home MCP Lets Claude, ChatGPT Control Devices

What happened

Google has begun rolling out early access to Home MCP, a Model Context Protocol server that lets outside AI agents — not just its own Gemini for Home — see and control devices across the Google Home ecosystem. The rollout was announced September 16, 2026, in a post by Taylor Lehman, group product manager for Google Home & Nest, and covered near-simultaneously by TechCrunch, The Verge, 9to5Google, Unite.AI, ByteIota and others 24101214. Google named Google Antigravity, Anthropic's Claude, Hermes and OpenClaw as launch-ready agents, and TechCrunch and Thurrott both report ChatGPT is also supported 1210.

The server exposes five capabilities documented on Google's own developer site: list_homes, list_home_resources, list_home_states, run_home_actions and list_home_history 11. In practice, that lets a connected agent enumerate rooms and devices, check whether a thermostat or Nest doorbell is online, issue commands, and pull historical event logs and camera activity — covering Nest hardware, Matter devices and anything labeled Works with Google Home 111213. Google and several outlets cite the same illustrative use cases: asking an agent what the kids did after arriving home from school and getting a cross-camera summary with clips, tallying how many loads of laundry ran in a week, or having the agent ping a Google Home speaker with a voice message once a task finishes 212141617.

Access is gated tightly at launch: US-only, English-only, and restricted to subscribers of Google Home Premium Advanced, the $20-per-month ($200-per-year) tier 2101516. Setup is developer-grade rather than one-click — creating a Google Cloud project, enabling the Home API, configuring OAuth consent and credentials, and pointing a compatible client such as Claude, Antigravity or OpenClaw at home.googleapis.com/mcp 111213. Google has also introduced a companion Home Developer MCP server that grounds coding assistants like Claude Code, Cursor and GitHub Copilot in Home API and Matter/Thread documentation, authenticated separately from the consumer-facing server 1216.

Why it matters

Home MCP does not replace Gemini for Home, which remains the fixed default assistant in the Google Home app and on Nest speakers 1415. What changes is that a household can now run a second, independent agent layer on top of Google's device and data plane. The Verge frames this as Google's real objective: positioning Google Home as the infrastructure layer beneath whichever agent wins on reasoning or personality, much as AWS became infrastructure for software companies rather than the application itself 14. That reading dovetails with the broader MCP wave documented elsewhere — Data Commons MCP opening public datasets to LLMs, X launching a hosted MCP server for its API, and ReleasePad building an MCP server so coding agents can draft changelogs 368 — all cases where a platform lets outside models act as the interface while the underlying company keeps the data and control plane.

MCP itself is the protocol layer that makes this interoperability possible without Google building bespoke integrations for every chatbot. As the Model Context Protocol's own documentation explains, servers expose tools (executable actions), resources (contextual data) and prompts (templates), while the connecting AI application decides how to use them 1819. That same standardization logic underpins Google's separate Agent2Agent (A2A) protocol, launched in April 2025 with more than 50 partners including Salesforce, SAP and Workday, which lets agents coordinate with each other rather than merely calling tools 20. Home MCP is a tool-and-data integration, not an A2A deployment, but it creates the physical-world endpoint a future multi-agent workflow — say, a household agent delegating an energy-optimization task to a specialist agent — could eventually use.

For enterprise AI, the episode functions as a case study, not a comparison. The same design questions Google had to answer for a family's living room — how much authority to grant a model, how to gate sensitive actions, how to log and revoke access — are the questions companies exposing CRM, ticketing or building-management systems to agents will face. A recent overhaul of the MCP specification itself has shifted more of that security burden onto developers and platform operators rather than the protocol 9, and a real-world case reinforces the stakes: security researchers found a backdoored npm package, postmark-mcp, silently exfiltrating emails from more than 1,600 downloads before it was caught 7.

Where the reporting agrees

Across TechCrunch, The Verge, 9to5Google, Unite.AI, ByteIota, Thurrott and Google's own documentation, the core facts are consistent. Home MCP launched in early access on September 16, 2026, restricted to US-based Google Home Premium Advanced subscribers at $20/month or $200/year 210111516. Claude, Antigravity, Hermes and OpenClaw are named as launch agents, with TechCrunch and Thurrott adding ChatGPT to that list 1210. Every outlet that describes functionality agrees on the same four capability buckets — structure discovery, state monitoring, device control and historical/event analysis — and nearly all cite the identical example of an agent summarizing what children did after school using footage from multiple cameras 21213141617. There is also unanimous agreement that Gemini for Home is not being replaced; it remains the default assistant in the Home app and on Nest speakers, with MCP functioning as an added layer rather than a swap 1415. Reporters and Google's documentation likewise agree that sensitive actions like unlocking doors are explicitly blocked, that automation creation is not yet supported, and that setup requires a Google Cloud project and OAuth configuration rather than a simple toggle 1112131617.

Where it doesn't

The discrepancies are mostly matters of framing and emphasis rather than contradicted facts. The Verge treats the launch as primarily a strategic infrastructure play for Google, arguing the company is content to lose the assistant war as long as it keeps the device and data layer, likening the ambition to AWS's role underneath internet businesses 14. TechCrunch's account is more product-focused and skeptical of near-term reach, stressing that Google declined to say whether or when the feature might expand beyond the Advanced tier, other countries or other languages 210. Cravingtech goes further than most in speculating about long-term consequences, suggesting users may eventually pick an AI agent the way they now pick a phone OS, independent of which smart-home brand they own 17 — a framing none of the more product-centric outlets (9to5Google, ByteIota) make as explicitly, since they concentrate on setup mechanics and available tools instead 413.

There is also a minor sourcing gap worth flagging: ChatGPT's inclusion as a supported agent appears in TechCrunch and Thurrott's coverage but is not listed among the named agents in Google's own developer documentation or in the Unite.AI and ByteIota accounts, which stick to Antigravity, Claude, Hermes and OpenClaw as the examples Google cited by name 1210111213. That's not necessarily a contradiction — MCP is designed so any compliant client can connect — but it means ChatGPT's specific participation rests on secondary reporting rather than Google's primary announcement.

The reading the evidence supports

The weight of overlapping, independently sourced detail — identical capability lists, identical pricing, identical use-case examples appearing in Google's own developer documentation as well as in TechCrunch, The Verge, Unite.AI and ByteIota's independent write-ups — makes clear this is a real, narrowly-scoped early-access feature, not vaporware or an overstated announcement. The Verge's infrastructure framing is the more analytically useful of the two dominant readings: Google previously opened Home APIs to reach hundreds of millions of connected devices, and Home MCP is a natural extension of that same platform logic applied to agentic AI rather than a defensive response to Gemini's competitive position. The caution embedded in nearly every account — rate limits, blocked door-unlocking, OAuth friction, an admitted risk of "unexpected or even undesired behavior" — signals that Google is treating this as a deliberate, security-conscious pilot, and the enterprise parallels to MCP's known weak points, like the postmark-mcp backdoor, suggest that caution is warranted rather than performative.