Google opens Google Home to outside AI agents via Model Context Protocol

Google has added Model Context Protocol integration to Google Home, letting third-party AI agents — Claude, Open Claw, Hermes, and Google Antigravity — reach directly into connected devices, event history, and the control layer for the first time. Taylor Lehman, group product manager for Google Home and Nest, wrote in the official blog that the connection turns the chosen agent into something that can "work securely with all devices and event history in the Google Home ecosystem."

MCP is a standard protocol originally developed by Anthropic that lets language models communicate with external tools and data sources in real time. Until now, Google smart-home control ran through Gemini for Home, the official interface that replaced Google Assistant in the Home app and on Nest speakers. The new integration does not replace Gemini; it adds a parallel path. The external agent plugs straight into the data and control layer, with no Google assistant in between. That means Claude or any MCP-compatible agent can correlate events across cameras — what the kid did when they got home from school — tally laundry cycles from device-state history, push a voice announcement through a Nest speaker when a task finishes, or build a custom control dashboard.

At launch, access is limited to Google Home Premium Advanced subscribers in the United States ($20 a month or $200 a year), rolling out gradually over the coming weeks. Setup requires creating a Google Cloud project and configuring it for Home MCP — not a plug-and-play flow for the average user, more for developers and power users willing to work in the cloud console. Lehman noted the protocol enforces rate limits and safety guards; for example, an agent cannot unlock smart locks. But he also warned that "connecting an agent to Home MCP may result in unexpected or undesired behavior," and urged developers to review Google's developer policies and terms of service.

Home Assistant, the open-source smart-home platform, already shipped a similar MCP integration. The author tested it last year with Claude for "vibe coding" the smart home — a term for writing automations and configs in natural language instead of YAML or graphical UIs. The agent handled tasks that standard interfaces (and users) struggle with: debugging, creating complex automations, designing dashboards, advanced configuration. The real shift isn't that Google Home can now flip lights via the Claude app; it's that Claude, or whatever agent you choose, now has access to the home's underlying data and control plane. That's a step toward a "whole-home computer" that understands what's happening, reasons across history, and acts on your behalf.

A system that can troubleshoot, suggest improvements, and spot patterns across the house is more useful than one that answers "sorry, I can't do that." But the gap between an impressive demo and a production-ready product with permissions for locks, thermostats, and HVAC systems remains wide. Google has put up guardrails, yet responsibility for the agent's behavior sits with its developer — and in practice, with the user who connects it. For now this is a tool for developers and early adopters; the open question is when, if ever, it reaches the mainstream user without a cloud project requirement.