deConz MCP

Máy chủ MCP cho máy chủ deConz do Dresden Elektronik phát triển (cầu nối giữa các nền tảng tự động hóa nhà thông minh và mạng Zigbee không dây)

GitHubDùng thử MCP nàyĐược tài trợ

Tài liệu

deConz MCP Server

An MCP (Model Context Protocol) server that exposes the deCONZ REST API to AI assistants. Control Zigbee lights, sensors, groups, scenes, rules, and schedules through natural language.

Supports stdio, SSE, and Streamable HTTP transports. HTTP transports are protected by a configurable bearer token.

Requirements

Python ≥ 3.10
uv (recommended) or pip
A running deCONZ / Phoscon gateway with a ConBee or RaspBee adapter
A valid deCONZ REST API key (see Obtaining an API key)

Installation

# Clone the repository
git clone https://github.com/marcinn2/deconz-mcp.git
cd deconz-mcp

# Install with uv (creates an isolated virtual environment)
uv sync

# Or install with pip into your environment
pip install -e .

Obtaining an API key

Open the Phoscon App in your browser (usually http://<gateway-ip>/pwa).
Go to Menu → Settings → Gateway → Advanced.
Click Authenticate app — this opens the network for 60 seconds.
Within those 60 seconds, run:

curl -s -X POST http://<gateway-ip>/api \
  -H "Content-Type: application/json" \
  -d '{"devicetype": "deconz-mcp"}'

The response contains your API key:

[{"success": {"username": "YOUR-API-KEY-HERE"}}]

Store it as DECONZ_API_KEY.

Quick start

stdio (Claude Desktop)

DECONZ_HOST=192.168.1.10 DECONZ_API_KEY=abc123def deconz-mcp

Streamable HTTP with bearer auth

DECONZ_HOST=192.168.1.10 \
DECONZ_API_KEY=abc123def \
MCP_AUTH_TOKEN=my-mcp-secret \
deconz-mcp --transport streamable-http --host 0.0.0.0 --port 8080

Combined SSE + Streamable HTTP

DECONZ_HOST=192.168.1.10 \
DECONZ_API_KEY=abc123def \
MCP_AUTH_TOKEN=my-mcp-secret \
deconz-mcp --transport server --host 0.0.0.0 --port 8080

Environment variables

Variable	Required	Default	Description
`DECONZ_HOST`	Yes*	—	IP or hostname of the deCONZ gateway
`DECONZ_PORT`	No	`80`	HTTP port of the deCONZ gateway
`DECONZ_API_KEY`	Yes*	—	deCONZ REST API key
`DECONZ_TLS`	No	`false`	Set `true` to use HTTPS
`DECONZ_TLS_VERIFY`	No	`true`	`false` accepts a self-signed gateway certificate; any other value is a path to a CA bundle
`DECONZ_ALLOW_RUNTIME_CONFIG`	No	see‡	Set `true` to expose `configure_deconz` even when the gateway comes from the environment
`MCP_AUTH_TOKEN`	No†	—	Bearer token clients must send to this MCP server
`MCP_BASE_URL`	No	`http://<host>:<port>`	Public base URL (used as OAuth issuer URL)

* Not required when using stdio and calling configure_deconz at runtime.
† Strongly recommended for HTTP transports exposed beyond localhost.
‡ Defaults to enabled only when DECONZ_HOST / DECONZ_API_KEY are absent. Once the gateway is configured from the environment, the tool is withheld so that a client cannot repoint this server at another host.

CLI reference

usage: deconz-mcp [--transport {stdio,sse,streamable-http,server}]
                  [--host HOST] [--port PORT]
                  [--log-level {DEBUG,INFO,WARNING,ERROR}]
                  [--auth-token TOKEN] [--base-url URL]

Flag	Default	Description
`--transport`	`stdio`	Transport mode
`--host`	`127.0.0.1`	Bind address (HTTP transports)
`--port`	`8000`	Listen port (HTTP transports)
`--log-level`	`INFO`	Logging verbosity
`--auth-token`	`$MCP_AUTH_TOKEN`	MCP bearer token
`--base-url`	`$MCP_BASE_URL`	OAuth issuer URL

Transport modes

Mode	Endpoint(s)	Description
`stdio`	stdin/stdout	Pipe-based — for Claude Desktop and local use
`sse`	`/sse`, `/messages/`	Legacy SSE (MCP pre-2025-03-26)
`streamable-http`	`/mcp`	Modern Streamable HTTP (MCP 2025-03-26)
`server`	all of the above	SSE + Streamable HTTP on one port

Authentication

deCONZ API key

The deCONZ REST API key is a gateway credential that travels as part of the URL path (/api/<apikey>/...). It is configured server-side via DECONZ_API_KEY and is never exposed to MCP clients: request URLs are kept out of tool errors, and the HTTP client's request logging is silenced because it would otherwise write the key to stderr on every call.

Resource IDs supplied by a client (light, group, scene, sensor, rule, schedule, and Touchlink IDs) are validated before they are placed in a request path, so a crafted ID cannot escape its resource and reach, for example, the gateway's configuration or its list of API keys.

MCP bearer token

The MCP_AUTH_TOKEN / --auth-token option protects the MCP server itself. Every HTTP request from an MCP client must include:

Authorization: Bearer <token>

The token is verified with a constant-time comparison to prevent timing attacks. When running over localhost only (default bind 127.0.0.1), bearer auth is optional but recommended.

Tools

Connection / configuration

Tool	Description
`configure_deconz`	Point the server at a deCONZ gateway at runtime (host, port, API key). Registered only when the gateway is not configured from the environment, or when `DECONZ_ALLOW_RUNTIME_CONFIG=true`
`get_gateway_config`	Read gateway name, firmware, Zigbee channel, IP, WebSocket port
`set_permit_join`	Open the Zigbee network for new device pairing (0–255 seconds)

Lights

Tool	Description
`list_lights`	List all lights with on/off, brightness, and reachability
`get_light`	Full JSON details for a single light
`set_light_state`	Control power, brightness, hue, saturation, colour temp, xy, effect, alert
`rename_light`	Rename a light
`delete_light`	Remove a light from the gateway

Groups

Tool	Description
`list_groups`	List all groups with member count and action state
`get_group`	Full JSON details for a single group
`create_group`	Create a new group (optionally pre-populate with lights)
`set_group_action`	Control all lights in a group simultaneously
`modify_group`	Rename a group or change its member lights
`delete_group`	Delete a group (lights remain)

Scenes

Tool	Description
`list_scenes`	List all scenes for a group
`get_scene`	Full JSON details for a scene
`create_scene`	Create a scene (captures current group state)
`recall_scene`	Activate a scene
`store_scene`	Overwrite a scene with the current group state
`rename_scene`	Rename a scene
`delete_scene`	Delete a scene

Sensors

Tool	Description
`list_sensors`	List all sensors with latest readings and battery levels
`get_sensor`	Full JSON details for a single sensor
`rename_sensor`	Rename a sensor
`set_sensor_config`	Update sensor config (enabled, battery level, sensitivity)
`delete_sensor`	Remove a sensor from the gateway

Rules (automations)

Tool	Description
`list_rules`	List all automation rules with status and trigger counts
`get_rule`	Full JSON details (conditions + actions) for a rule
`create_rule`	Create a new rule with conditions and actions
`set_rule_status`	Enable or disable a rule
`delete_rule`	Delete a rule

Schedules

Tool	Description
`list_schedules`	List all timed schedules
`get_schedule`	Full JSON details for a schedule
`create_schedule`	Create a new schedule with ISO 8601 time expression
`set_schedule_status`	Enable or disable a schedule
`delete_schedule`	Delete a schedule

Touchlink

Tool	Description
`touchlink_scan`	Start a Touchlink scan (~10 s) to find nearby Zigbee devices
`get_touchlink_results`	Return results from the last Touchlink scan
`touchlink_identify`	Make a Touchlink device blink for identification
`touchlink_reset`	Factory-reset a Touchlink device

Resources

Resources are read-only, cacheable snapshots that MCP clients can fetch without issuing tool calls.

URI	Description
`deconz://config`	Gateway configuration JSON
`deconz://lights`	All lights with state JSON
`deconz://groups`	All groups with action state JSON
`deconz://sensors`	All sensors with state JSON
`deconz://rules`	All automation rules JSON
`deconz://schedules`	All schedules JSON
`deconz://state`	Complete gateway state (all resources combined)

Prompts

Prompts are pre-built conversation starters that guide the AI through common workflows.

Prompt	Arguments	Description
`home_overview`	—	Full status report: all lights, sensors, reachability, anomalies
`control_lights`	`room` (optional)	Turn lights on/off, set brightness or colour
`manage_scenes`	`group_id` (optional)	Create, recall, update, or delete scenes
`add_device`	—	Step-by-step guide to pair a new Zigbee device
`setup_automation`	—	Create a rule triggered by a sensor event
`diagnose_device`	`device_name` (optional)	Diagnose unreachable or misbehaving devices
`evening_routine`	`bedtime` (default `23:00`)	Activate an evening scene and schedule lights off

Claude Desktop configuration

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "deconz": {
      "command": "deconz-mcp",
      "env": {
        "DECONZ_HOST": "192.168.1.10",
        "DECONZ_API_KEY": "your-api-key-here"
      }
    }
  }
}

Or if installed in a virtual environment:

{
  "mcpServers": {
    "deconz": {
      "command": "/path/to/deconz-mcp/.venv/bin/deconz-mcp",
      "env": {
        "DECONZ_HOST": "192.168.1.10",
        "DECONZ_API_KEY": "your-api-key-here"
      }
    }
  }
}

HTTP client configuration

For Streamable HTTP transport:

{
  "mcpServers": {
    "deconz": {
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer my-mcp-secret"
      }
    }
  }
}

For legacy SSE transport:

{
  "mcpServers": {
    "deconz": {
      "url": "http://localhost:8080/sse",
      "headers": {
        "Authorization": "Bearer my-mcp-secret"
      }
    }
  }
}

Docker

A multi-platform image (linux/amd64 + linux/arm64) is published to GitHub Container Registry by the release workflow whenever a version tag is pushed:

ghcr.io/marcinn2/deconz-mcp:latest
ghcr.io/marcinn2/deconz-mcp:0.2
ghcr.io/marcinn2/deconz-mcp:0.2.1

A multi-stage Dockerfile is also included if you prefer to build locally. The builder stage uses the official uv image to install dependencies and compile the package as a wheel; the runtime stage is python:3.14-slim and runs as the unprivileged user app (uid 1000).

Pull

docker pull ghcr.io/marcinn2/deconz-mcp:latest

Build locally

# Single-platform (current machine)
docker build -t deconz-mcp:latest .

# Multi-platform push (requires a buildx builder with multi-platform support).
# Releases do this automatically; this is for pushing to another registry by hand.
docker buildx build \
  --platform linux/amd64,linux/arm64 \
  --tag registry.mne.pl/deconz-mcp:latest \
  --tag registry.mne.pl/deconz-mcp:0.2.1 \
  --push .

Run — HTTP server (SSE + Streamable HTTP)

docker run -p 8000:8000 \
  -e DECONZ_HOST=192.168.1.10 \
  -e DECONZ_API_KEY=abc123def \
  -e MCP_AUTH_TOKEN=my-mcp-secret \
  ghcr.io/marcinn2/deconz-mcp:latest

Run — stdio

docker run -i \
  -e DECONZ_HOST=192.168.1.10 \
  -e DECONZ_API_KEY=abc123def \
  ghcr.io/marcinn2/deconz-mcp:latest --transport stdio

Docker Compose

A ready-to-use Compose file is in docs/docker-compose.yml. It defines two services:

Service	Transport	Started by default
`deconz-mcp`	`server` (SSE + Streamable HTTP) on port 8000	Yes
`deconz-mcp-stdio`	`stdio`	No — requires `--profile stdio`

# Copy and edit the environment file
cp .env.example .env   # set DECONZ_HOST, DECONZ_API_KEY, MCP_AUTH_TOKEN

# Start the HTTP server (--env-file is required: the compose file lives in docs/,
# so a repository-root .env is not picked up automatically)
docker compose --env-file .env -f docs/docker-compose.yml up

# Run a one-shot stdio session
docker compose --env-file .env -f docs/docker-compose.yml --profile stdio run --rm deconz-mcp-stdio

Use the stdio service in Claude Desktop:

{
  "mcpServers": {
    "deconz": {
      "command": "docker",
      "args": ["compose", "-f", "/path/to/docs/docker-compose.yml",
               "--profile", "stdio", "run", "--rm", "deconz-mcp-stdio"],
      "env": {
        "DECONZ_HOST": "192.168.1.10",
        "DECONZ_API_KEY": "your-api-key-here"
      }
    }
  }
}

Kubernetes

The manifest at docs/k8s-deployment.yaml contains all resources needed to run the server in a cluster:

Resource	Purpose
`Namespace`	`deconz-mcp` — isolates all resources
`Secret`	`DECONZ_API_KEY` and `MCP_AUTH_TOKEN` (base64-encoded)
`ConfigMap`	`DECONZ_HOST`, `DECONZ_PORT`, `DECONZ_TLS`, `MCP_BASE_URL`
`Deployment`	1 replica, non-root, read-only root FS, resource limits
`Service`	`ClusterIP` on port 80 → pod 8000
`Ingress`	Commented-out template for nginx / cert-manager

Deploy

# 1. Encode your secrets
echo -n 'your-api-key'   | base64   # → paste into Secret.DECONZ_API_KEY
echo -n 'your-mcp-token' | base64   # → paste into Secret.MCP_AUTH_TOKEN

# 2. Edit the ConfigMap (DECONZ_HOST, MCP_BASE_URL) in docs/k8s-deployment.yaml

# 3. Apply
kubectl apply -f docs/k8s-deployment.yaml

# 4. Verify
kubectl -n deconz-mcp get pods
kubectl -n deconz-mcp logs -f deploy/deconz-mcp

Health check

kubectl -n deconz-mcp port-forward svc/deconz-mcp 8000:80
curl http://localhost:8000/health
# {"status": "ok", "deconz_configured": true}

The Deployment configures both a liveness probe and a readiness probe against /health, so Kubernetes automatically restarts the pod if the server becomes unresponsive.

Ingress (optional)

Uncomment the Ingress section at the bottom of docs/k8s-deployment.yaml and set your hostname. Do not add a rewrite-target annotation: the endpoints live at /mcp, /sse and /messages/ and must reach the pod with their paths intact. TLS termination happens at the ingress controller; the pod always speaks plain HTTP internally.

Health check

HTTP transports expose a liveness probe:

curl http://localhost:8080/health
# {"status": "ok", "deconz_configured": true}

Project structure

deConz-mcp/
├── Dockerfile                      # Multi-stage image build (runs as uid 1000)
├── pyproject.toml                  # Package metadata and dependencies
├── uv.lock                         # Locked dependency versions (committed)
├── .env.example                    # Template for docker compose
├── README.md
├── docs/
│   ├── docker-compose.yml          # Compose services (HTTP + stdio)
│   └── k8s-deployment.yaml         # Kubernetes: Namespace, Secret, ConfigMap,
│                                   #   Deployment, Service, Ingress (template)
├── tests/
│   └── test_smoke.py               # Client, credential-handling and MCP surface tests
└── src/
    └── deconz_mcp/
        ├── __init__.py
        ├── __main__.py             # CLI entrypoint and transport wiring
        ├── client.py               # Async deCONZ REST API HTTP client
        └── server.py               # FastMCP server: tools, resources, prompts

deCONZ API overview

The server covers these API categories:

Category	Endpoints
Config	`GET /config`, `PUT /config` (permit join)
Lights	`GET /lights`, `GET /lights/<id>`, `PUT /lights/<id>/state`, `DELETE /lights/<id>`
Groups	`GET /groups`, `POST /groups`, `PUT /groups/<id>/action`, `DELETE /groups/<id>`
Scenes	Full CRUD under `/groups/<id>/scenes/<sid>` incl. recall and store
Sensors	`GET /sensors`, `PUT /sensors/<id>/config`, `DELETE /sensors/<id>`
Rules	Full CRUD under `/rules/<id>`
Schedules	Full CRUD under `/schedules/<id>`
Touchlink	`POST /touchlink/scan`, identify, reset

For the full API reference, see the deCONZ REST API documentation.

Data & Privacy

Preliminary assessment only — not legal advice. See full notes below.

Personal home use

When this server runs in a private household and is accessed only by the residents, processing of smart-home device data is likely covered by the household exemption (GDPR Recital 18). In that scenario the GDPR does not apply and no additional compliance steps are required.

Commercial or shared deployments

Deploying this server in offices, rental properties, hotels, co-working spaces, or any environment where you process data on behalf of other people takes you outside the household exemption. In those cases:

Presence and motion sensor data constitutes personal behavioral data (Art. 4(1) GDPR). Establish a documented lawful basis (Art. 6) before processing it.
Conduct a Data Protection Impact Assessment (Art. 35) if the deployment involves systematic monitoring of occupants on a large scale.
Provide a privacy notice to data subjects describing what is collected, for how long, and under what legal basis.

Security recommendations

Risk	Recommendation
API key travels in URL paths, so the gateway's own access logs record it	Rotate the deCONZ API key periodically; restrict access to gateway logs. This server keeps the key out of its own logs and out of the errors it returns to clients
Unencrypted transport	Enable TLS for any network-facing deployment (`DECONZ_TLS=true`); use a reverse proxy with a valid certificate. Set `DECONZ_TLS_VERIFY=false` only for a self-signed gateway certificate on a trusted network
MCP endpoint publicly accessible	Always set `MCP_AUTH_TOKEN` when binding to a non-loopback address

What this software does NOT do

No data is sent to third parties, analytics services, or cloud providers.
No telemetry, tracking pixels, or consent libraries are present in this codebase.
All communication stays between the MCP client, this server, and the local deCONZ gateway.

Legal notice

This GDPR assessment was generated as a preliminary, exploratory evaluation. It does not constitute legal advice and does not replace a legal audit. For binding guidance, consult a qualified data protection lawyer in your jurisdiction.

Releasing

Releases are cut by pushing a version tag. .github/workflows/release.yml then:

Checks that the tag matches the version in pyproject.toml and fails early if it does not.
Runs the full CI workflow (lint, format, type-check, tests on Python 3.10–3.12).
Builds the image for linux/amd64 and linux/arm64, with an SBOM and build provenance, and pushes it to ghcr.io/marcinn2/deconz-mcp.
Creates the GitHub release with generated notes, the image digest, and the built wheel and sdist attached.

# bump version = "0.3.0" in pyproject.toml first, then:
git tag -a v0.3.0 -m "v0.3.0"
git push origin v0.3.0

A tag containing a hyphen, such as v0.3.0-rc.1, is published as a pre-release and does not move the latest image tag.

Image tags produced for v0.3.0: 0.3.0, 0.3, and latest.

No secrets need to be configured: the workflow authenticates to GHCR with the built-in GITHUB_TOKEN. After the first release, set the package visibility under Packages → deconz-mcp → Package settings if the image should be pullable anonymously.

License

Apache 2.0 — see LICENSE.

Disclaimer

This software is not affiliated with or endorsed by Dresden Elektronik. Use at your own risk. It does not come with any warranty of any kind. There is no liability for the developer. This software is a personal project that I maintain in my free time. Refer to the licence for more information.