Someone who wants their agent to skim the camera roll or dim the hallway lights can now route that through Google’s own plumbing. The company published a Model Context Protocol server for its home ecosystem on Tuesday and opened early access.
Reach is narrow at the start. Only US subscribers of Google Home Premium Advanced, the $20-a-month tier with 60 days of event video history, can connect. Google said any client able to call MCP tools qualifies and named Claude, ChatGPT, Hermes, OpenClaw and its own Antigravity suite. Four areas of capability are exposed: enumerating rooms and devices, reading live state, executing device actions and mining past events.
Connecting is a project rather than a switch. The owner spins up a Google Cloud project, turns on the Home API, registers an OAuth client, then hands the server address to the agent, which prompts for permission. Google published walkthroughs for three clients.
Once wired, the assistant has real context about a house. Google’s own examples include summarising an afternoon across several camera feeds, tallying laundry loads over a week, or playing a spoken announcement when a long background task finishes.
Limits are spelled out. Doors cannot be unlocked by an agent, rate limits apply, and familiar-face data needs the property manager’s separate consent. Google warns that a live home can behave unexpectedly under an agent and recommends a test property, and access can be revoked from the Home app at any time. Some device traits remain experimental.
