Google hands five outside coding agents access to its Home devices

Daniel Bader, who writes 9to5Google's Inbox newsletter, wrote that he wants Home MCP for one job: "I'm totally going to use it to teach my Nest Cam Outdoor that the cat who wanders onto my driveway every morning isn't, in fact, a person." ^[8] He already runs MCP against his bank, his to-do app TickTick and his Obsidian database ^[6]. This ships for someone who has already wired an agent into his own systems and now wants the house in the same chat window.

Two conditions gate it. A Google Home Premium Advanced subscription and a US account ^[2], plus one of the coding agents Google names, of which there are five: Claude, Codex, Antigravity, Hermes and OpenClaw ^[1][9]. Both conditions have to hold at once, so the eligible population is an overlap of a paid tier and a habit, not the Google Home installed base ^[10].

The described scope splits cleanly in two. Google said MCP "gives your favorite AI agent real-world physical context" ^[3], and lists four uses: to "analyze and answer questions about home activity", "summarize camera footage across rooms", "manage devices on your behalf", and "create your own smart home dashboards using everyday language" ^[4]. Three of those four read the home, and the fourth changes it. The newsletter says nothing about what permissions a connected agent receives, or how a user revokes them ^[12].

This is Google's second answer to the same problem in the same year. Gemini for Home was the company's own assistant layer, rolled out to make automation more straightforward ^[5]. Home MCP hands the devices to somebody else's agent instead, and Bader wrote that he "commend[s] the company for dipping its toe into the DIY space and seeing what happens" ^[7].

For anyone deciding what to connect on Monday, two axes separate the candidate actions: whether you can undo the result, and whether you would find out it happened without going to look. Camera summaries and activity questions are undoable and visible, because the output is text you read. A device that got switched at two in the morning is undoable and invisible until you wake up cold. Third-party agents fail in the second corner more expensively than the first, and the read-only half also carries its own cost: summarising footage across rooms means a household's video record passes through an outside vendor's context.

I'd hold the scope to the read-only half of Google's own list, which is questions about home activity and camera summaries ^[4]. Device management moves across only when there is a log that shows which agent did what. The tradeoff is real, because the reason to connect an agent at all is that it can act, and a read-only connection cannot act on devices.