On September 16, 2026, Google began early access to Home MCP, an official connection framework for Google Home. AI agents capable of calling MCP tools—such as Claude and OpenClaw—can now discover devices in a home, read their states and history, and execute authorized actions. Smart home management, previously restricted to the Google Home app or Gemini for Home, is now being extended to agents of the user's choosing. While this brings new convenience, it also means AI decisions now reach into the physical world—lights, cameras, climate control—so the freedom to choose a connection destination inherently raises new questions of trust.
Breaking Down the Home Into Five Tools
According to official documentation, the capabilities Home MCP exposes to agents fall into five categories. list_homes enumerates accessible homes, while list_home_resources returns devices, rooms, device traits, and available commands. list_home_states reads connection status and current values, and run_home_actions sends commands to target devices. The final tool, list_home_history, traces state changes and event records over a specified period.
This combination allows agents to handle commands like "turn off all the outdoor lights" as well as answer questions such as "what happened while I was away?" using historical data. Google has demonstrated examples including summarizing a child's activity after coming home by cross-referencing footage from multiple cameras, and counting how many times a washing machine ran over a week based on its state history. Other envisioned uses include voice notifications via Google Home speakers when an asynchronous task completes, or building custom dashboards for a household.
However, these are examples from early access users, not specifications guaranteeing accuracy or universal device compatibility. Google's phrase "all devices and event history" is itself constrained by what each device exposes, user permissions, subscription plans, and how much history is retained. MCP provides a common calling convention, but it cannot create capabilities that don't already exist on the device side.
Google's MCP Offerings Are Not a Single Front Door
The only thing the four MCP lines have in common is the connection method itself. Home MCP handles home states, history, and device control; Home Developer MCP handles technical documentation; Workspace MCP handles business data and app operations; and Google Cloud MCP handles cloud resources.
| Line | Primary Scope | What Agents Do | Authentication & Release Stage |
|---|---|---|---|
| Home MCP | Google Home devices, rooms, states, event history | State queries, history analysis, authorized device control | User OAuth, Early Access |
| Home Developer MCP | Google Home developer docs, Matter 1.5.1, Thread 1.4.1, OpenThread | Searches technical documentation for development answers | Unauthenticated shared tier or API key |
| Google Workspace MCP | Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat, People | Reading, drafting, file operations, scheduling | OAuth 2.0, Developer Preview |
| Google Cloud MCP | BigQuery, Cloud Storage, Cloud Run, Compute Engine, IAM, etc. | Queries and manages cloud resources via product-specific tools | Google Cloud authentication; GA or Preview depending on product |
The comparison axis here is the resources and actions described in Google's official documentation—not a comparison of tool counts or performance across servers. Home Developer MCP, despite its similar name, is not a connection point that operates real devices; rather, it lets development-support AI search Google Home and Matter specifications.
The difference also shows up in what kind of accident can occur. A mistaken email draft created in Workspace and lights or climate control activating in the middle of the night via Home MCP have very different real-world consequences, even though both stem from the same kind of MCP call. Google Cloud is centered on organizational management boundaries like IAM and compute resources. Home MCP, by contrast, must factor household co-owners and physical devices into its permission design.
A Ban on Unlocking Doesn't Fully Explain the Safety Design
Google has implemented rate limits on Home MCP and prohibited sensitive operations such as unlocking doors. But in the same warning section, the company also states plainly that depending on the connected agent, unexpected or undesirable behavior can occur. Having safeguards in place is not the same as guaranteeing that an agent's judgment will be safe.
To connect, users must enable the Home API within a Google Cloud project and create an OAuth client. Users then select a home and grant permissions on the authentication screen. Access can be revoked at any time through the Google Home app or a Google Account. When connecting to a shared home, Google requires that other household members be informed that an agent can operate devices and access home data.
Camera-related data carries an even finer boundary. Using facial recognition data for registered individuals through Home MCP requires a separate, explicit consent—one that only a home's administrator can grant. This also presumes the presence of a compatible Nest camera or doorbell with facial recognition enabled on each device. Google Home's policy for third-party apps likewise requires businesses seeking access to sensitive devices, such as locks, to pass certification confirming secure data handling and restricted use.
What emerges is a design that doesn't end once authentication is granted. The action itself is banned for sensitive cases, call frequency is throttled, additional consent is required for sensitive data, and users retain the ability to sever the connection afterward. The remaining vulnerability lies in cases where an agent misreads context while operating entirely within its authorized permissions.
Early Access: US, English, $20 a Month
At launch, Home MCP is available only to US-based Google Home Premium Advanced subscribers, and only in English. The Advanced plan in the US costs $20 per month or $200 per year. Using Home MCP also requires a functioning Google Home setup, a Google Cloud project, and an MCP-compatible client—this is not a feature the average user can simply switch on from a settings screen.
Google's guidance lists Antigravity, Claude Cowork, and OpenClaw as example connections. After registering an OAuth redirect URI and configuring a client ID and secret, users select the target home from the Google Home permissions screen. Connection destinations and authorization scopes must be registered following Google's setup instructions.
In Japan, Google Home Premium Advanced is sold for ¥2,000 per month or ¥20,000 per year, but subscribing to that plan alone does not grant access to Home MCP. The official announcement limits the rollout to the US and English language, with no indication of a timeline for Japan. According to TechCrunch, Google did not respond to questions about when it might expand to other pricing plans or markets.
Whether Home MCP takes hold will depend less on the number of compatible agents and more on how effectively misoperations within authorized permissions can be minimized during early access. If users can clearly distinguish the scope of control for each device, household co-owners remain aware of active connections, and problematic connections can be revoked immediately, then home AI could shift from being a closed feature tied to one specific assistant into an open connection framework that users are free to reconfigure.