On September 16, Google began rolling out early access to a Google Home MCP server — a standard interface that lets outside AI agents read from and act on the devices in your home. Not Gemini-only: Google's own docs walk you through connecting Claude and OpenClaw alongside its own Antigravity, and any agent that speaks the Model Context Protocol can plug in.

For a company that spent a decade building a walled garden around Nest, this is a real change of posture. Let's look at what actually shipped, and why the shape of it matters more than the headline.

What it does

Once connected, an agent can see and act on the devices and event history across the Google Home ecosystem — Nest devices, Matter devices, and anything wearing the "Works with Google Home" badge. Google's own examples: ask it what happened in the house today, get a summary of camera footage across rooms, hand it routine device management, or describe a dashboard in plain language and have it built for you.

The gate is narrow for now. It's US-only, it requires an active Google Home Premium Advanced subscription ($20/month on its own), and setup means standing up a Google Cloud project and granting MCP permissions yourself. It's early access for paying subscribers — with a developer-shaped setup — not a mass consumer rollout yet.

The door has a frame

Here's the detail we liked most: Google ships the integration with rate limits and a hard exclusion list. Third-party agents cannot perform sensitive actions like unlocking doors. Full stop.

That's good product design, and it's the part most teams get wrong when they bolt an agent onto a system. The instinct is to expose everything and let the model "decide." Google did the opposite: the platform decides which capabilities are agent-safe, and the boundary lives in the platform, not in the prompt. A prompt can be argued with. An API that doesn't have an unlock method cannot.

Why MCP is the real story

Model Context Protocol has quietly become the USB-C of AI integration — a single connector that lets any client talk to any tool. When Google, of all companies, chooses an open protocol over a proprietary Gemini-only hook, it tells you where the market has landed: the agent is the user's choice, and platforms compete on being good to integrate with.

If you run a product with an API, that's your signal. Your customers' agents are coming whether you build for them or not. The question is whether they arrive through a door you designed — with a frame, a lock list, and rate limits — or through whatever scraped, brittle path they can find.

What we'd tell a client

  • Publish an MCP surface for the 20% of your product that is safe and useful to automate. Start small.
  • Decide the exclusion list before the capability list. What must never happen without a human? Write that down first.
  • Treat rate limits as UX. An agent that can hammer your system will, eventually, by accident.

Google just published the reference implementation of a careful agent integration. Copy the posture, not just the feature.