Google Opens Home Controls to Outside AI Agents
Google opened early access to its Model Context Protocol (MCP) server for the Google Home ecosystem on September 16, 2026. TechCrunch The server lets third-party AI agents work directly with smart home devices and event history through that standard connection.
MCP is a common standard that lets assistants call outside tools without custom code for each one. At launch, the supported agents are Claude, ChatGPT, Hermes, OpenClaw, and Google Antigravity. All are MCP-compatible, so they can use the Home tools with no separate integration per assistant. Access is gated. Control is delegated.
Google positions the server for natural-language operation. Documented uses include reviewing camera summaries, monitoring smart home activity, controlling connected devices, and building custom smart home dashboards. Underneath, that is two basic functions: asking about past device events, and acting on current device state, both in conversation.
Device coverage follows the normal Google Home compatibility rules. That includes Google Nest doorbells and thermostats, plus Works with Google Home or Matter devices such as light bulbs. Google describes no separate certification for MCP access. If a device already works in Home, it falls within the controllable surface.
Setup is not one-click. An administrator must create a Google Cloud project, configure it for Home MCP, pass the MCP setup details to the chosen agent, then sign in and grant permissions. The Cloud project acts as the control point for policy. Permissions are granted by the user at link time.
Early access started September 16, 2026 and continues over the coming weeks for Google Home Premium Advanced subscribers in the U.S. Premium Advanced costs $20 per month. Google declined to comment on if or when the server would extend to other subscription tiers or markets. Google is asking early adopters for feedback through its Smart Home for Developers Community.
Separately, Google Home Developers maintains an MCP Servers page at developers.home.google.com. It documents a developer-focused MCP server meant to help AI agents and tools give accurate technical answers about Google Home developer programs and technology. A July 29 specification notes that this developer server connects to several development environments in either unauthenticated or authenticated mode.
The broader context here is a change in where smart home decisions happen. Voice routines and app automations keep the reasoning inside one vendor stack. An MCP server flips that arrangement. Device state, history, and controls become tools any compatible assistant can call, with the model putting the steps together.
In my view, the details worth watching are permission granularity and history retention. Camera summaries and activity monitoring are useful because they condense long event streams, but they raise scoping questions a light switch does not. Which devices, which time ranges, and which household members an agent can query will matter more for shared homes than total device coverage. The Cloud project plus sign-in flow suggests Google wants those controls stated clearly from the start.
Looking at what this enables, custom dashboards may matter as much as voice-style control. If an agent can read live state and history and then build a view, the dashboard stops being a fixed app screen and becomes something made on demand. It is a familiar pattern from cloud monitoring, now applied to the home. The long arc is positive: open tool interfaces tend to produce more lasting automation than closed routines, if the trust model holds.
