Google has begun opening Google Home to outside AI agents that can check connected devices, read household event history and issue commands. The early-access feature, called Home MCP, could let a compatible assistant review camera activity or turn off the outside lights through a conversation. It also gives that assistant access to records of what happens at home.
Google’s September 16 release notes announce the rollout. According to TechCrunch’s Sarah Perez, Google said access would reach US Google Home Premium Advanced subscribers over the following weeks. That plan costs $20 a month or $200 a year. Connecting an agent requires developer setup, so this is currently an option for enthusiasts comfortable configuring software.
Model Context Protocol, or MCP, gives an AI application a standard way to call tools in another service. After Google authorization for a selected home, Home MCP lets the connected client retrieve permitted household data and request supported device actions. The client’s own approval settings determine when it asks before using those tools. Disconnecting stops future access, but does not necessarily erase information already received by the client or its provider.
The useful possibility is asking an assistant what happened while you were away and having it consult actual home records. Google’s history reference describes timestamps, device names, state changes and events, potentially including attribution to a user, agent or application. Camera events can also return optional media URLs when requested. That could reveal household routines as well as whether a device switched on.
Those are documented response fields, not results from a household we tested. Available history depends on device support, permissions and retained subscription history. Camera-event media access also does not establish access to continuous video footage.
| Access | What an authorized agent can receive or do | Confirmed boundary |
|---|---|---|
| Discover the home | Retrieve home and device information, potentially including address, time zone, room groupings, equipment details and supported commands. | Returned information depends on the authorized home and exposed resources. |
| Check current conditions | Read device connectivity and supported states. | Device support and user permissions constrain access. |
| Read history | Query past events and state changes, with optional camera-event media URLs. | Available records depend on device support, permissions and retained history. Familiar-face data requires separate consent. |
| Control devices | Send supported commands with parameters to a target device. | Google says it blocks sensitive actions such as unlocking doors. Creating or managing automations is not supported. |
Google’s resource reference and action reference make an important distinction: discovering a device does not give an agent every imaginable control over it. Available commands depend on the capabilities that device exposes and the user’s grant.
The setup guide describes authorization through OAuth, the sign-in process used to grant an application access to an account’s resources. Users configure the client, authenticate in a browser and select a home. The documented flow was not independently observed for this article.
Familiar-face information needs a separate consent link, approved by a manager of the home. It also requires a compatible Nest camera or doorbell with familiar-face detection enabled. That extra permission covers familiar-face data, not all camera events or household records.
For everyone else living there, Google tells the person making the connection to inform household members that the agent can control devices and access home data. The guide does not document an approval requirement for every resident, guest or person represented in those records. One person’s connection can therefore affect more than their own privacy.
Permission to connect is also different from approval for each action. Google Antigravity defaults unconfigured MCP tools to asking, but allows broader rules that let tools run without a prompt. Claude Cowork’s permission modes likewise vary whether connector reads and writes need approval; its Auto mode makes its own risk decision rather than guaranteeing a human checkpoint. These are documented client behaviors, not observed Home MCP tests.
Google says Home MCP applies rate limits and prohibits sensitive actions such as unlocking doors. Its guide nevertheless warns of “unexpected or even undesired behavior.” Those protections constrain what an agent can request; they do not establish that it will choose the right permitted action. The guide supplies neither rate-limit values nor a complete list of prohibited actions, and the safeguards have not been independently tested here.
Users can revoke access through the Google Home app or Google Account. Google’s general guidance for connected applications says stopping access may still leave previously shared data with the outside developer, whom users may need to contact for deletion. Whether home records are retained depends on the chosen client, provider, plan and settings.
Getting this far requires a Google Cloud project, enabling the Home API, creating and publishing an external OAuth application, configuring its callback address and entering credentials in the AI client. Google also lists access approval without explaining its criteria or timing. It declined to tell TechCrunch whether or when other plans or countries would become eligible.
We would start with Google’s suggested separate test home before connecting a shared household. The useful experiment is whether an assistant can reliably answer questions and carry out bounded requests. Giving it access to everyone’s home records deserves a separate decision from whether it can turn off a light.