Google opened early access to a Model Context Protocol server for Google Home on September 16, 2026, so agents like Claude and ChatGPT can control connected devices and read camera summaries in plain language.
- This guide provides comprehensive, actionable information
- Consider your specific workflow needs when evaluating options
- Explore our curated AI Assistants tools for specific recommendations
What actually shipped
On September 16, 2026, Google gave early access to a new Model Context Protocol (MCP) server for Google Home, letting AI agents control connected devices, read back camera summaries, and pull smart-home activity history using natural language instead of the Google Home app's UI. It's one of the first mainstream smart-home platforms to hand an AI agent direct access to a home's cameras and connected devices, reported first by TechCrunch.
What it can actually touch
The MCP server covers Google Nest doorbells, Nest thermostats, any device certified "Works with Google Home," and Matter-standard devices — the article's own example is a Matter light bulb. Beyond direct control, it can review camera-generated activity summaries and surface smart-home event history, which is the part that goes further than a voice command ever did: an agent can be asked what happened at the front door today, not just told to lock it.
If you're building rather than just asking, Google also says the same access can power custom smart-home dashboards — the MCP connection is the same one either way, just driven by your own code instead of a chat window.
Which agents actually work with it
The launch names Claude, ChatGPT, Google Antigravity, and two others (Hermes, OpenClaw) as compatible at launch. This is MCP doing what it's for: one server, multiple agent clients, rather than Google building a separate integration per assistant. For background on why that architecture exists at all, see agent tool architecture and MCP.
Setup is a Google Cloud project, not a toggle
This isn't a switch you flip in the Google Home app. Getting an agent connected means: creating a Google Cloud project, configuring it for Home MCP, handing the resulting MCP configuration to whichever agent you're using, then completing a sign-in and permissions grant that the agent itself initiates. Google's setup guide lives in the Google Home Developer Center. That's a materially higher bar than turning on a skill — this is aimed at developers and technically comfortable early adopters first, not a mass-market feature yet.
What Google didn't say
Two gaps are worth flagging rather than glossing over. First, the article states plainly that Google wouldn't comment on if or when the MCP would roll out more broadly — not to other subscription tiers, not to other countries. Treat the $20/month US-only scope as the whole offering for now, not a starting point with a known roadmap. Second, on security, Google's own language is that agents can "securely work with" connected devices — the reporting doesn't include specifics on encryption, permission scoping, or what an agent can see versus control. For a feature that reaches home security cameras, that's a thin amount of detail to hand a decision on, and worth asking about directly if you're the one setting it up for a household.
Who should actually set this up now
For the wider agent landscape this plugs into, see Google Antigravity and every AI assistant in the directory. For the concept underneath this launch, agent tool architecture and MCP covers what MCP actually is and why it beats a one-off API integration.
Explore curated tools related to this guide: