Netatmo Energy

Netatmo Energy MCP

English | Français

[CI] [License: Apache-2.0] [Node.js >= 22.19]

A Netatmo MCP server for Netatmo smart thermostats and smart radiator valves. It gives AI assistants structured access to your heating:

• room temperatures and setpoints
• heating demand and boiler activity
• temperature history
• deterministic heating analytics
• weekly heating schedules
• optional heating control: room setpoints, away / frost-guard mode and schedules, with every change confirmed by you

It runs on your machine and talks only to the official Netatmo Energy API. It is read-only by default: it can change your heating only if you enable write mode at login.

It works with any Model Context Protocol client that runs local servers, whatever the model:

• Claude: Claude Desktop, Claude Code
• GPT / OpenAI: Codex CLI, VS Code + GitHub Copilot, Cursor
• Gemini: Gemini CLI, VS Code + GitHub Copilot
• Mistral and others through multi-model clients
• Local LLMs (Llama, Qwen, Mistral, DeepSeek, …): LM Studio, and Ollama via Goose, Continue, Cline, AnythingLLM, LibreChat or Open WebUI

It also works in Windsurf / Devin Desktop, Zed, Roo Code, Kilo Code, JetBrains AI Assistant, Kiro and Warp. See all compatible clients.

Status: early release (0.2.0). The Netatmo API integration has been validated on a real installation; feedback and device reports are welcome.

Why this project?

A Netatmo heating system records useful data:

• the temperature and setpoint of every room
• what each radiator valve is asking for
• when the boiler is asked to heat

That data is locked in the Netatmo app, where you can look at it but not ask questions about it.

This project is a small bridge between the Netatmo Energy API and any MCP-compatible AI assistant. Ask in plain language: "Which room was coldest last night?" The assistant calls a precise tool and answers from your data. It doesn't guess. With write mode enabled, you can also say "Set the bedroom to 19 °C until 7 am", then confirm the change.

The few other Netatmo MCP servers target Netatmo weather stations. This one is built for thermostats, radiator valves and heating history. See docs/research.md for the comparison.

Features

• Discovery: homes, rooms, thermostats, smart radiator valves, relays and gateways.
• Current status
  • Per room: temperature, setpoint, setpoint mode and heating demand.
  • Boiler on/off and open-window detection.
  • Device health: battery, radio/Wi-Fi signal, reachability.
• History
  • Room temperature and setpoint history at 30 min to 1 month resolution.
  • Boiler activity history (installations with a Netatmo thermostat).
  • Long ranges are fetched in chunks and summarised, so answers stay small.
  • Gaps are reported, never filled in.
• Heating analytics. These are deterministic calculations; no AI model computes the numbers.
  • Per-room statistics.
  • Time below, within or above the setpoint.
  • Largest temperature drop and cool-down rates.
  • Room rankings.
  • Rule-based detection of unusual readings, each with severity and confidence.
• Weekly schedules: zones (Comfort, Night, Eco…), room setpoints per zone and the timetable, as readable days and times.
• Heating control (opt-in write mode)
  • Temporary room setpoints that always end (3 h by default, at most 24 h), or a return to the schedule.
  • Home mode: schedule, away or frost guard, optionally until a date.
  • Switch, create, edit and rename weekly schedules.
  • Every change is previewed and needs your explicit confirmation. Temperatures are limited to 7–28 °C by default.
• 15 read-only MCP tools, 6 heating-control tools, 4 resources and 4 prompts (daily report, anomaly review, room comparison, heating pattern review).
• Simple setup
  • OAuth2 browser login with a single login command.
  • Tokens stored securely and refreshed automatically.
  • doctor checks your setup.

Quick start

You need Node.js 22.19 or later and a Netatmo account with Energy devices.

1. Create a free Netatmo developer app

1. At https://dev.netatmo.com/apps, choose Create.
2. Set the redirect URI to http://localhost:8977/callback.
3. Keep the client ID and client secret for step 2.

Step-by-step guide: docs/authentication.md.

2. Log in

npx -y netatmo-energy-mcp login

login asks for the client ID and secret, then opens your browser so you can sign in on netatmo.com and approve read-only access. To let the assistant change your heating, run login --write instead (see write mode). Then check the setup:

npx -y netatmo-energy-mcp doctor

Running from source instead: clone the repository, run pnpm install && pnpm build, then use node /path/to/netatmo-energy-mcp/dist/index.js in place of npx -y netatmo-energy-mcp.

3. Connect your AI assistant

Most clients use the same mcpServers JSON block. This works for Claude Desktop, Cursor, Windsurf / Devin Desktop, Cline, Roo Code, Kiro, LM Studio, JetBrains AI Assistant, AnythingLLM, Warp and Gemini CLI:

{  "mcpServers": {    "netatmo-energy": {      "command": "npx",      "args": ["-y", "netatmo-energy-mcp"]    }  }}

Command-line clients:

# Claude Codeclaude mcp add --transport stdio --scope user netatmo-energy -- npx -y netatmo-energy-mcp# OpenAI Codex CLIcodex mcp add netatmo-energy -- npx -y netatmo-energy-mcp# GitHub Copilot CLIcopilot mcp add netatmo-energy -- npx -y netatmo-energy-mcp

VS Code + GitHub Copilot. Add this to .vscode/mcp.json; note the servers key:

{  "servers": {    "netatmo-energy": { "type": "stdio", "command": "npx", "args": ["-y", "netatmo-energy-mcp"] }  }}

No secrets go in any of these files: login stored them in your user configuration folder. File locations for each client, plus Zed, Continue, Goose, LibreChat, Kilo Code, Msty and Open WebUI, are in examples/.

4. Ask

"What's the temperature in each room right now?"

Example questions

These are the kinds of questions the tools are built for. The assistant chooses the tools; the table shows which ones answer each question.

Boiler "run time" is the time the thermostat requested heat. Netatmo does not measure gas or energy consumption, so this project never reports it.

Available MCP tools

The read-only tools need only the read_thermostat OAuth scope. The full reference, with arguments and outputs, is generated from the server itself: docs/tools.md.

Write mode only. Each change needs your confirmation.

Resources: netatmo://homes and netatmo://homes/{homeId}/rooms, …/devices and …/status.

Write mode (heating control)

Write mode is off by default. To enable it, log in again with:

npx -y netatmo-energy-mcp login --write

This also requests the write_thermostat OAuth scope. The heating control tools appear after you restart your MCP client.

Safeguards:

• You confirm every change. Clients that support MCP elicitation ask you directly. With other clients, the first call only returns a preview and a one-time token. The assistant must show you the preview and get your agreement before calling again. Nothing is sent to Netatmo before that. This second flow relies on the assistant following its instructions; set NETATMO_MCP_CONFIRM=elicitation to allow changes only through a confirmation prompt shown by your client. If your client cancels every change without showing anything, set NETATMO_MCP_CONFIRM=token.
• Limits. Temperatures must stay between 7 and 28 °C. Manual setpoints end after 3 h by default and 24 h at most. Change these with NETATMO_MCP_MIN_TEMP, NETATMO_MCP_MAX_TEMP and NETATMO_MCP_MAX_SETPOINT_HOURS.
• Kill switch. NETATMO_MCP_WRITE=0 forces read-only mode, even with a write-enabled login.
• Audit log. Every applied or failed change is appended to changes.log in your configuration folder.
• No automatic retries. A failed write is never resent blindly.

Netatmo has no API to delete a schedule: schedules created here can only be deleted in the Netatmo app. Renaming a schedule and choosing a schedule with home mode "schedule" use undocumented Netatmo parameters, so they are marked experimental. Details: docs/configuration.md.

Supported devices

Run netatmo-energy-mcp probe and open a device compatibility report to help extend this table. The probe output is sanitized.

Compatible AI assistants and MCP clients

This is a standard local (stdio) MCP server, so it is not tied to one AI vendor. Any MCP client that can start local servers can use it, with whatever model that client runs.

Configuration for each client, checked against official documentation: examples/.

Not compatible: assistants that only accept remote MCP servers (ChatGPT apps/connectors, Claude.ai on the web, Mistral Le Chat). This server runs locally by design, so your credentials never leave your machine.

Testing. The server is tested with the official MCP SDK client and the MCP Inspector. Tool-calling quality with small local models varies by model. Please report any client-specific problem.

Authentication

This project uses Netatmo's OAuth2 authorization code flow. You sign in on netatmo.com; your Netatmo password is never seen by this tool.

Scope. By default only read_thermostat is requested, so a leaked token couldn't change your heating. login --write also requests write_thermostat.

Your own app. Each user registers a free Netatmo developer app. The client secret cannot be shipped in open-source code.

Refresh. Access tokens are refreshed automatically. Several MCP clients running at once coordinate through a lock file, so they don't invalidate each other's tokens.

Details: docs/authentication.md.

Privacy & security

What leaves your machine. Only HTTPS requests to api.netatmo.com. In read-only mode they only reach 4 read endpoints: the client refuses any write before it touches the network, and tests enforce this. In write mode, a change is sent only after you confirm it.

What stays local.

• Credentials live in credentials.json in your user configuration folder, readable only by your account (0600 on macOS/Linux; a restricted ACL on Windows).
• No telemetry, no analytics and no third-party services.
• The server talks to your MCP client over stdio and opens no network port.

Your AI provider. Data that tools return is read by the assistant you use, so it is processed by that model provider.

No account data or location. Your email and home coordinates are never returned.

Read more:

• Security model
• Pre-release security review
• Reporting a vulnerability

Architecture

[Architecture: an AI assistant talks over stdio to the local server, which calls the Netatmo API (read-only by default)]

The Netatmo client and the analytics are independent of MCP. Design decisions are recorded as ADRs, and docs/architecture.md describes the layers.

Limitations

These come from the Netatmo Energy API; details are in docs/api-capabilities.md.

• No energy or gas consumption. Boiler activity is heat-demand time. With aggregated data, the number of burner cycles cannot be known.
• No history of valve heating demand. It is only available as a current value. A future opt-in collector may record it (ADR-0011).
• No outdoor temperature in the Energy API. Weather context is planned for v0.4.
• History resolution is 30 minutes at best. Each request returns at most 1024 values, so long ranges use coarser scales.
• The documentation doesn't match the API in places. Live data shows that boiler measures are in seconds, not the documented minutes. This project converts them.
• Rate limits. Netatmo limits requests per user and per app. The server throttles itself and caches topology and current status.

Roadmap

Write mode will never be enabled by default. See docs/roadmap.md.

Contributing

Bug reports, device compatibility reports and pull requests are welcome. See CONTRIBUTING.md to set up the project, which runs on mocked API responses: no Netatmo account needed. Please follow the Code of Conduct.

License

Apache-2.0

Disclaimer

This is an independent open-source project. It is not affiliated with, endorsed by or sponsored by Netatmo or Legrand. "Netatmo" is a trademark of its owner and is used here only to describe compatibility. Use at your own risk; heating safety must never depend on an AI assistant.