Google just opened its smart home to AI agents. On September 16, 2026, the company launched early access to a Model Context Protocol (MCP) server for Google Home, allowing Claude, ChatGPT, and other AI agents to control connected devices, review camera summaries, and build custom dashboards using natural language.
This is the first mass-market moment for MCP beyond ChatGPT plugins. It signals a shift: smart home devices are no longer just hardware features. They're becoming software layers that AI agents can orchestrate.
For founders building automation products, this changes the economics. You can now hand off device control to an agent instead of building custom integrations. But you're also competing with Google's own MCP server, which means your value has to live elsewhere.
What Changed: MCP as the Smart Home Interface
Google Home has existed since 2016. You could ask Google Assistant to turn on lights, set thermostats, check cameras. That was single-device control, one command at a time.
The MCP server removes the Assistant middleman. Instead of talking to Google, you talk to your AI agent. The agent then uses MCP to:
- Control any Google Nest device (Nest Hub, Nest Mini, Nest Doorbell, Nest Thermostat, Nest Camera)
- Control any "Works with Google Home" device (Matter, Philips Hue, LIFX, etc.)
- Review camera summaries and motion alerts
- Access smart home event history
- Build custom automation dashboards The difference: agents can now reason about your home. Not "turn on the living room light," but "it's 6 PM and the sun's setting—do I need porch lights on? Check if anyone's home. If not, turn them off. If someone is home, set them to 30% brightness."
That's orchestration, not commands.
Real Specs: Devices, Access, Rollout
Supported Devices: - All Google Nest hardware (Nest Hub Max, Nest Mini, Nest Audio, Nest Doorbell, Nest Thermostat, Nest Camera) - Any device certified as "Works with Google Home" - Matter devices (any Matter-certified smart home device) - Estimated coverage: 500+ device models across 100+ brands Supported Agents: - Claude (Anthropic) - ChatGPT (OpenAI) - Hermes (LM Studio) - OpenClaw (open-source) - Google Antigravity (internal) Access & Pricing: - Rollout: Starting today, September 16, 2026 - Limited to: Google Home Premium Advanced subscribers in the US ($20/month) - Timeline: Rolling out over the coming weeks - Future: Google hasn't committed to broader availability (other tiers, other regions) Setup: - Create a Google Cloud project - Configure MCP for Home via the Developer Center - Sign in and grant permissions - Agent then controls your devices Authentication: - OAuth 2.0 (standard web auth) - Granular permissions per device category - All commands logged in Google Home event history
Use Cases: When Agents Touch Hardware
MCP for Google Home creates three practical use cases for founders and home automation:
1. Task Delegation (Non-Technical) - "Schedule my bedroom lights to mimic sunrise at 6 AM for the next 30 days" - "If motion is detected at 2 AM, turn on hallway lights at 10% brightness" - "Turn off all lights except living room if no one's home" Founders building smart home UIs can now let agents handle the rule creation. Instead of building a rules engine, you delegate to Claude.
2. Multi-Device Orchestration - HVAC + humidity control (thermostat adjusts if humidity crosses threshold) - Energy optimization (check solar production, load-shift to high-production hours) - Security automation (arm system, lock doors, record if motion detected while away) Agents excel at this because they can reason across devices. A thermostat alone can't optimize for energy. An agent can.
3. Reactive Home Assistants - Morning routine (lights → temperature → coffee maker → news summary) - Departure (lock doors → arm system → close shades → save energy) - Arrival (unlock door → turn on lights → adjust temperature → check security) These routines are hardcoded today. With agents, they become responsive. "I'm arriving home with groceries—unlock the door and turn on kitchen lights" instead of "turn on kitchen lights."
Cost Breakdown: Premium Tier is the Barrier
Entry cost: - Google Home Premium Advanced: $20/month ($240/year) - Requires: Compatible Google Home device ($99–$399) - Total year-one: $240–$600 depending on hardware For context: - Standard Google Home (free tier): No MCP access - Google Home Premium Basic: No MCP access (exact tier features unclear) - Google Home Premium Advanced: Only tier with MCP access Cost comparison to alternatives: | Option | Monthly | Control | Reasoning | Automation | |--------|---------|---------|-----------|------------| | Google Home Premium Advanced + Claude | $20 | Native | Yes | Yes | | Home Assistant + HA Yellow hub | $9.99 | Open | No | Rules-based | | SmartThings Premium | $9.99 | Samsung only | No | Rules-based | | Apple Home | Included | Apple only | No | Automations |
The barrier: $20/month is high for early adopters. For a founder testing product-market fit, that's not trivial. But for a user with 10+ smart home devices, it's cheaper than paying for three separate automation services.
When Google Home MCP Wins vs Loses
Google Home MCP wins when: - You have 5+ Google Home devices (at that scale, $20/month makes sense) - You want agent-based reasoning (Claude can explain why it chose an action) - You want privacy-aware control (everything stays on your Google account, logged) - You want to delegate device management to a non-technical person (natural language) - You're already a Google Workspace customer (single sign-on, unified logging) Google Home MCP loses when: - You have fewer than 3 connected devices ($20/month is expensive) - You have non-Google devices (Matter support exists, but limited) - You want full local control (Google Cloud authentication required) - You need custom logic Home Assistant can't handle (stuck with agent's reasoning) - You're privacy-sensitive about cloud logging (all events stay in Google account) Honest tension: Google owns both the device ecosystem and the control layer. Competitors building smart home agents can use MCP, but Google can outprice you by integrating MCP for free in the future. Your differentiation has to be the agent itself (better reasoning, multimodal, custom integrations), not the device control.
The Pattern: MCP Becomes the Interface Standard
This isn't just about Google Home. It's about MCP becoming the interface standard for hardware → agent interaction.
Six months ago, MCP was ChatGPT-only. Today: - Claude Desktop has MCP servers (Gmail, Google Drive, GitHub, etc.) - Google Workspace supports MCP - Google Cloud supports MCP - Now Google Home supports MCP Next: Amazon Alexa, Apple Home, Samsung SmartThings will likely support MCP. Not this year, maybe not next. But the pattern is clear.
Founders building smart home products have two paths: 1. Build on MCP: Let agents handle logic, you focus on device integrations 2. Build the agent: Compete with Claude/ChatGPT on reasoning, use their MCP to touch hardware Path one is commodity work. Path two is where the value moves.
Practical Framework: Should You Care?
If you're building automation software: - Does your user have a Google Home device? (30% of US homes do) - Are they willing to pay $20/month? (Only if they have 5+ devices) - Is your value in orchestration or explanation? (If explanation, agents beat you) If you're building a smart home device: - Can you expose control via MCP? (Yes, Matter devices already do) - Does MCP make your device more valuable? (Yes, agents can discover it) - Are you competing with Google's own integration? (Yes, always) If you're building with Claude: - Can you give users device control? (Yes, now you can via MCP) - Is natural language control enough? (Only if users accept agent decisions) - Do you need real-time device state? (MCP supports this, but may have latency) For founders: MCP is infrastructure. Your opportunity is in the agent layer (better reasoning) or the integration layer (devices Google doesn't support). Device control itself is becoming too cheap to sell.
Bitroot helps founders design systems for AI-first automation. Explore founder guides