zigbee2mqtt-mcp

A Model Context Protocol server for administering a Zigbee2MQTT estate.

Most Zigbee integrations already let an assistant turn a light on. This one is for the layer underneath: mesh health, weak links, devices that keep rejoining, stale batteries, firmware updates, pairing, binding, reporting intervals and device options — the things you normally open the Zigbee2MQTT frontend for.

It talks directly to the Zigbee2MQTT MQTT bridge API, so it works regardless of whether you use Home Assistant, Node-RED, openHAB or nothing at all. Tested against Zigbee2MQTT 2.x; several tools depend on bridge endpoints that only exist there, including bridge/health, device/binds/clear and the OTA scheduling topics.

Why this exists

If you already run Home Assistant, your Zigbee devices are exposed there and an assistant can control them. What Home Assistant does not expose is the bridge itself: link quality, mesh topology, interview state, OTA availability, permit_join, device options, bindings and reporting configuration. This server fills exactly that gap.

Design

• No database. Zigbee2MQTT publishes its bridge topics as retained messages, so a fresh subscription yields a complete picture in a few hundred milliseconds. The server keeps only the latest payload per topic in memory.
• No daemon. Pure stdio. It starts and dies with the MCP session.
• Correlated requests. Every bridge/request/* carries a transaction id and is matched to its bridge/response/*, so concurrent calls cannot cross wires.
• Tiered writes. Destructive operations are gated behind a write mode and an explicit confirm argument.

Install

Requires Node.js 22 or newer and network access to the MQTT broker that Zigbee2MQTT uses.

npx zigbee2mqtt-mcp

VS Code / GitHub Copilot

Add to your MCP configuration:

{
  "servers": {
    "zigbee2mqtt": {
      "command": "npx",
      "args": ["-y", "zigbee2mqtt-mcp"],
      "env": {
        "Z2M_MQTT_URL": "mqtt://192.168.1.10:1883",
        "Z2M_MQTT_USERNAME": "mqtt",
        "Z2M_MQTT_PASSWORD": "${input:z2mPassword}"
      }
    }
  }
}

Claude Desktop

{
  "mcpServers": {
    "zigbee2mqtt": {
      "command": "npx",
      "args": ["-y", "zigbee2mqtt-mcp"],
      "env": {
        "Z2M_MQTT_URL": "mqtt://192.168.1.10:1883"
      }
    }
  }
}

Point Z2M_MQTT_URL at your MQTT broker, not at the Zigbee2MQTT frontend port. If Zigbee2MQTT's configuration.yaml says mqtt.server: mqtt://192.168.1.10:1883, use that value verbatim.

Behind a corporate npm proxy? If npm config get registry is not https://registry.npmjs.org/, your proxy's upstream feed may not carry this package. Add an explicit override to the args rather than changing your global registry: "args": ["-y", "--registry", "https://registry.npmjs.org/", "zigbee2mqtt-mcp"]

Configuration

Write modes

Tools above the active tier are not registered at all, so a model cannot reach for them.

Irreversible tools (z2m_remove_device, z2m_ota_update, z2m_touchlink, z2m_restart_bridge, z2m_set_bridge_options) additionally require confirm: true on every call. Three safe tools reduce state rather than only adding to it, so they are annotated destructive: z2m_rename_device, z2m_manage_group and z2m_bind. Their reducing actions also require confirm: true — z2m_manage_group (remove, remove_all_members) and z2m_bind (clear). z2m_rename_device is annotated destructive because it breaks anything referencing the old name, but stays unconfirmed since renaming back restores it.

Every tool also advertises MCP tool annotations so a client can decide what may run without prompting:

Annotations describe the tool, not the tier: a safe tool can still be destructive. The tier decides what is registered, the annotation tells the client what to ask about.

Opt in to the destructive tier only when you want it:

"env": { "Z2M_WRITE_MODE": "full" }

Tools

Read

Safe writes

z2m_check_updates, z2m_permit_join, z2m_set_device_options, z2m_rename_device, z2m_configure_device, z2m_interview_device, z2m_set_state, z2m_manage_group, z2m_bind, z2m_configure_reporting

Full writes

z2m_remove_device, z2m_ota_update, z2m_restart_bridge, z2m_set_bridge_options, z2m_touchlink

Data availability caveats

Some fields depend on your Zigbee2MQTT configuration. The server detects what is available and tells you rather than silently returning nothing.

Security notes

• The broker URL may carry credentials (mqtt://user:pass@host). They are redacted everywhere the URL is echoed back to a model or written to a log, and error text from the MQTT client is scrubbed before it leaves the process. Zigbee2MQTT logs its own broker URL at startup, so lines relayed by z2m_get_logs are scrubbed too.
• safe mode can actuate anything on the mesh via z2m_set_state, including locks, valves and sirens. Use off if that matters.
• z2m_permit_join opens the network to any nearby Zigbee device for the duration of the window. It is in safe because it is routine and reversible, but it is a security boundary.

Example prompts

• "Is my Zigbee network healthy?"
• "Which devices have the weakest signal?"
• "Which batteries need replacing?"
• "Any firmware updates available?"
• "Open the network for pairing via the kitchen router for two minutes."
• "This sensor stopped reporting temperature — fix it."
• "Bind the hallway remote to the hallway light so it works if the bridge is down."

Development

npm install
npm run build
npm run watch

# Unit tests. No broker required; they run on every push across Node 22 and 24.
npm test