ccu-mcp

Talk to your HomeMatic smart home from Claude, Cursor, or any MCP client.

ccu-mcp connects to the CCU's built-in JSON-RPC API and exposes your devices, rooms, programs, and system variables as MCP tools. No addons, no XML-API, no cloud — just a direct connection to the CCU on your local network.

Works with any HomeMatic CCU: debmatic (HomeMatic on Debian), a CCU3, or OpenCCU (formerly RaspberryMatic) — anything that exposes the standard /api/homematic.cgi endpoint.

What can it do?

Ask your AI assistant things like:

• "What's the temperature in the bathroom?"

• "Are any windows open?"

• "Set the living room heating to 21 degrees"

• "Show me all devices with low battery"

• "What's the gas meter reading?"

• "Which devices have low battery or haven't been seen in a long time?"

• "Find all channels whose names don't match their device name"

• "Rename all devices to follow a consistent naming convention with floor labels (UG/OG/EG)"

• "Which room is the window sensor in?"

The MCP server handles device discovery, type resolution, session management, and value conversion — the AI just calls the tools.

Related MCP server: Home Assistant MCP

Prerequisites

• A running HomeMatic CCU (debmatic, CCU3, or OpenCCU — formerly RaspberryMatic) reachable on your network

• The CCU's admin username and password (the same credentials you use to log into the WebUI)

• Node.js 24+ (for running from source or stdio mode) or Docker

Quick start

export CCU_HOST=your-ccu-hostname-or-ip
export CCU_PASSWORD=your-ccu-admin-password
npx ccu-mcp --stdio

If it prints server_ready to stderr, it's working. Press Ctrl+C to stop. Now set it up in your MCP client — see below.

Installation

There are two ways to run this: stdio (the server runs as a subprocess of your MCP client) or HTTP (the server runs standalone in Docker and clients connect over the network). Pick one.

Option A: stdio (direct, simplest)

This is the easiest setup. Your MCP client (Claude Code, Cursor, etc.) starts the server as a child process — no Docker, no network config, no auth tokens.

For Claude Code, create a .mcp.json file in your project directory (or any directory where you'll use Claude Code):

{
  "mcpServers": {
    "ccu-mcp": {
      "command": "npx",
      "args": ["ccu-mcp", "--stdio"],
      "env": {
        "CCU_HOST": "your-ccu-hostname-or-ip",
        "CCU_PASSWORD": "your-ccu-admin-password"
      }
    }
  }
}

Replace your-ccu-hostname-or-ip with your CCU's hostname (like homematic-ccu3) or IP (like 192.168.1.50), and your-ccu-admin-password with the password you use to log into the CCU WebUI.

Restart Claude Code. Run /mcp to check it connected. You should see ccu-mcp in the list.

Alternatively, use the Claude Code CLI:

claude mcp add ccu-mcp -- npx ccu-mcp --stdio

Option B: Docker (standalone HTTP server)

Use this if you want the server running independently — for example on a home server, accessible to multiple clients, or when your MCP client supports HTTP remotes.

1. Start the container:

git clone https://github.com/claymore666/ccu-mcp.git && cd ccu-mcp
docker build -t ccu-mcp .
docker run -d \
  --name ccu-mcp \
  -e CCU_HOST=your-ccu-hostname-or-ip \
  -e CCU_PASSWORD=your-ccu-admin-password \
  -e MCP_ALLOWED_HOSTS=your-server-ip:3000 \
  -v ccu-data:/data \
  -p 3000:3000 \
  ccu-mcp

MCP_ALLOWED_HOSTS is required for remote clients. The server's DNS-rebinding protection rejects any request whose Host header isn't on the allowlist — by default only localhost/127.0.0.1/[::1] on the MCP port. Set it to every name/IP clients will use to reach the server (comma-separated, host:port). Without it, the local health check works but every remote MCP request gets 403 Invalid Host header.

2. Get the auth token. The server generates a random bearer token on first startup and saves it inside the container's data volume. You need this token to authenticate your MCP client. Grab it with:

docker exec ccu-mcp grep MCP_AUTH_TOKEN /data/.env

This prints something like MCP_AUTH_TOKEN=e96suzi1iG0H-GPif6K2.... The part after = is your token.

3. Configure your MCP client. If your client uses .mcp.json, add the HTTP server:

{
  "mcpServers": {
    "ccu-mcp": {
      "url": "http://your-server-ip:3000",
      "headers": {
        "Authorization": "Bearer PASTE-YOUR-TOKEN-HERE"
      }
    }
  }
}

To inject the token automatically (requires jq):

TOKEN=$(docker exec ccu-mcp grep MCP_AUTH_TOKEN /data/.env | cut -d= -f2)
jq --arg t "$TOKEN" '.mcpServers["ccu-mcp"].headers.Authorization = "Bearer " + $t' .mcp.json > .mcp.json.tmp && mv .mcp.json.tmp .mcp.json

This only updates the ccu-mcp entry — other servers in your .mcp.json are left alone.

4. Check it's healthy:

curl http://localhost:3000/health

Browser-based clients (CORS)

By default the HTTP server sends no CORS headers, so a random web page can't drive a local instance. To let browser-based MCP clients like MCP Inspector connect directly, set MCP_ALLOWED_ORIGINS to a comma-separated allowlist of trusted origins (e.g. https://app.example,http://localhost:6274). A request whose Origin is on the list gets that exact origin reflected in Access-Control-Allow-Origin — never the wildcard *, which would let any site drive a local instance that controls real CCU hardware. A request from any other origin gets no CORS headers (the browser blocks it) and is rejected server-side by DNS-rebinding protection. Authentication is always enforced regardless: every MCP request needs the bearer token.

The HTTP transport also has DNS-rebinding protection on by default: it rejects requests whose Host header isn't localhost/127.0.0.1/[::1] on the configured port. If you reach the server under another hostname or IP (reverse proxy, container DNS name, the server's LAN address), list those hosts in MCP_ALLOWED_HOSTS or legitimate requests get a 403.

TLS. The bearer token travels in the request, so anything beyond loopback should be encrypted. You have two options: terminate TLS at a reverse proxy (Caddy/nginx) in front and bind the server to loopback (MCP_HOST=127.0.0.1), or let the server serve HTTPS itself by setting MCP_TLS_CERT and MCP_TLS_KEY to a PEM cert/key pair. Plain HTTP is still fully supported — it stays the zero-config default — but the server logs a warning at startup when it's serving the token over unencrypted HTTP on a non-loopback bind; set MCP_ALLOW_PLAINTEXT=true to acknowledge that and silence it.

Token rotation & expiry. By default the bearer token lives forever. Two optional, composable controls let you rotate it without dropping clients:

• Auto-generated token — set MCP_AUTH_TOKEN_TTL_DAYS (fractional days allowed) to give the generated token a lifetime. The server rotates it automatically at runtime shortly before it lapses (no restart needed; also on startup if it expired while the server was down), prints the new token on stderr, and keeps the just-replaced token validating for MCP_AUTH_TOKEN_GRACE_HOURS (default 24) so in-flight clients survive the swap. To force a rotation sooner, delete $CACHE_DIR/.env (or just its MCP_AUTH_TOKEN line) and restart.

• Explicit token — when you set MCP_AUTH_TOKEN yourself, you own its lifetime (TTL doesn't apply). To rotate, put the new token in MCP_AUTH_TOKEN, move the old one to MCP_AUTH_TOKEN_PREVIOUS, and restart; both are accepted during the overlap. Drop MCP_AUTH_TOKEN_PREVIOUS and restart once every client is on the new token. Comparison stays timing-safe across every currently-valid token.

Brute-force protection (fail2ban). The auto-generated token is 256 bits of randomness, so guessing it is infeasible. If you set MCP_AUTH_TOKEN yourself, make it long and random (e.g. openssl rand -base64 32) — a short or guessable token is the one case brute force matters. The server does not rate-limit or lock out failed logins in-process; that job belongs to a firewall-level tool like fail2ban, which bans the source IP before the request ever reaches the server. To make that easy, every rejected request logs a structured line to stderr:

{"ts":"2026-06-18T17:28:00.370Z","level":"warn","msg":"auth_failed","client":"203.0.113.7","hadToken":true}

Ready-to-use fail2ban config ships in fail2ban/: copy filter.d/ccu-mcp.conf to /etc/fail2ban/filter.d/ and the jail in jail.d/ccu-mcp.local to /etc/fail2ban/jail.d/ (it defaults to 5 failures in 10 minutes → 1-hour ban). The server logs to stderr, so point fail2ban at wherever you collect it — the journal (backend = systemd) when run as a unit, or a file when you redirect stderr/docker logs; both are spelled out in the jail file. Requires LOG_LEVEL=warn or lower (info, the default, is fine; error suppresses the line). Behind a reverse proxy the logged IP is the proxy's, so run fail2ban against the proxy's access log instead.

CORS support was first implemented by @marcinn2 in his fork marcinn2/debmatic-mcp — thanks!

HTTPS

If your CCU uses HTTPS (self-signed certificates are fine), add these environment variables:

CCU_HTTPS=true
CCU_PORT=443

The server accepts self-signed certificates automatically — certificate verification is off by default because CCUs ship with self-signed certs (the server logs a warning when running unverified). To actually verify the connection and close the MITM gap, you have three options:

• Pin the fingerprint (simplest for a self-signed appliance cert): set CCU_TLS_FINGERPRINT to the cert's SHA-256 (hex, with or without colons). The connection is rejected unless the CCU presents exactly that certificate. Read it with:

  echo | openssl s_client -connect "$CCU_HOST:443" 2>/dev/null | openssl x509 -noout -fingerprint -sha256

• Trust a CA/self-signed PEM: point CCU_CA_CERT at the certificate file for standard chain validation.

• System trust store: if your CCU has a publicly-trusted certificate, set CCU_TLS_VERIFY=true.

CCU_TLS_FINGERPRINT takes precedence over CCU_CA_CERT, which takes precedence over CCU_TLS_VERIFY.

Configuration

All configuration is via environment variables:

How to supply these (inline, .env, or export)

The required CCU_HOST / CCU_PASSWORD (and everything else) are environment variables. Provide them in whichever of these you prefer — you need just one:

• Inline in .mcp.json — the env block shown in Option A above. Simplest; self-contained.

• Shell export — as in Quick start above.

• A .env file — keeps secrets out of .mcp.json. The server does not read .env on its own, so load it with Node's built-in flag (Node ≥ 20.6):

  {
    "mcpServers": {
      "ccu-mcp": {
        "command": "node",
        "args": ["--env-file=/path/to/.env", "/path/to/ccu-mcp/dist/index.js", "--stdio"]
      }
    }
  }

  Copy .env.example to .env and fill it in (it documents every variable). Docker users can pass the same file with docker run --env-file .env or compose's env_file:. Keep .env gitignored.

Multiple CCU targets (profiles)

By default the CCU_* vars above configure a single CCU. To reach several CCUs (e.g. prod + dev) from one server, define named profiles instead. Set these the same way as any other config (inline, .env, or export — see above):

CCU_PROFILES=prod,dev
CCU_DEFAULT_PROFILE=prod           # active at startup (defaults to the first listed)

CCU_PROD_HOST=ccu.example
CCU_PROD_USER=ai
CCU_PROD_PASSWORD=...
CCU_PROD_HTTPS=true
CCU_PROD_PROTECTED=true            # writes need confirm:true

CCU_DEV_HOST=127.0.0.1
CCU_DEV_PORT=18080
CCU_DEV_USER=Admin
CCU_DEV_PASSWORD=                  # may be empty (e.g. an OpenCCU dev box)

Each profile takes the same settings as the flat vars, prefixed CCU_<NAME>_ (name upper-cased, non-alphanumerics → _): HOST (required), PASSWORD (may be empty), USER, PORT, HTTPS, TIMEOUT, SCRIPT_TIMEOUT, TLS_FINGERPRINT, CA_CERT, TLS_VERIFY — plus two policy flags:

• CCU_<NAME>_PROTECTED=true — write tools refuse unless called with confirm: true, which unlocks writes to that target for the rest of the session. Exception: run_script and delete_system_variable require confirm: true on every call — they never ride on the session unlock (scripts bypass all typed-tool guards; deletion is unrecoverable), and confirming them does not unlock the session for other writes.

• CCU_<NAME>_READONLY=true — write tools are refused outright.

With CCU_PROFILES unset, the flat CCU_* vars are used as a single default profile (unchanged behavior). At runtime, list_ccu_targets shows the targets, get_connection_info reports the active one, and use_ccu switches it. Read tools also accept an optional target to read from another CCU for a single call without switching.

Tools

28 tools organized by what you'd actually want to do:

Find things — list_devices, list_rooms, list_functions, list_interfaces, list_programs, list_system_variables, list_links, describe_device_type

Read state — get_value, get_values (bulk), get_paramset

Change things — set_value, put_paramset, set_system_variable, create_system_variable, delete_system_variable, assign_channel, unassign_channel, execute_program

Check health — get_service_messages, acknowledge_service_messages, get_rssi, get_system_info

Switch targets — list_ccu_targets, get_connection_info, use_ccu (multi-CCU profiles; see above)

Other — help (context-aware), run_script (raw HomeMatic Script for bulk operations, renaming devices/channels, querying room membership, or anything not covered by the other tools)

Most tools auto-resolve the interface and value types from the device address — you don't need to know whether a device is on BidCos-RF or HmIP-RF.

Resources and prompts

Besides tools, the server exposes MCP resources — browsable JSON snapshots your client can attach as context:

homematic://devices, homematic://rooms, homematic://functions, homematic://programs, homematic://sysvars, homematic://interfaces, homematic://device-types, homematic://system

The server polls the CCU in the background (every RESOURCE_POLL_INTERVAL seconds) and sends notifications/resources/updated for resources whose content changed — to clients that subscribed to them via resources/subscribe.

It also ships MCP prompts — ready-made workflows you can invoke from clients that support them (e.g. as slash commands in Claude Code):

• check-windows — are any windows or doors open?

• room-status — full status report for one room

• set-heating — set a room's target temperature

• good-night — prepare the house for night

• diagnostics — check for device issues

• device-info — detailed info about a device's capabilities and parameters

How it works

The server talks to the CCU's JSON-RPC API (the same one the WebUI uses). On startup it:

1. Logs in and caches the session (reused across restarts)

2. Loads the device type cache from disk (or warms it in the background)

3. Starts the MCP server on stdio or HTTP

Device type schemas are cached locally so the AI can look up valid parameters, types, and value ranges without hitting the CCU every time.

Values come back as native types — 21.5 not "21.500000", true not "true".

Tested devices

This has been tested against a production debmatic installation with:

• HmIP-eTRV-2 / eTRV-2 I9F (radiator thermostats)

• HmIP-STHD (wall thermostats with humidity)

• HmIP-WTH-2 (wall thermostats)

• HmIP-SWDO-I (door/window contacts)

• HmIP-STHO (outdoor temperature/humidity)

• HmIP-ESI (energy/gas meter)

• HmIP-FALMOT-C12 (floor heating controller)

• HmIP-HEATING (virtual heating groups)

• HmIP-WRCC2 (wall remote)

• HM-PB-6-WM55 (BidCos 6-button remote)

• RPI-RF-MOD (radio module)

Other device types should work too — the server queries the CCU for parameter descriptions rather than maintaining a static device database.

Changelog

Release notes — including behavior changes to check before upgrading (stricter config validation, /health response shape, per-session write confirmation, retry semantics) — live in CHANGELOG.md.

Related projects

• OpenCCU — community-maintained, cloud-free CCU firmware for Raspberry Pi, x86/ARM, and CCU3/ELV-Charly hardware (formerly RaspberryMatic; built on the OCCU framework)

• debmatic — Run HomeMatic on Debian, Ubuntu, Raspberry Pi OS, Armbian

• OCCU — eQ-3's original Open CCU SDK (the upstream HomeMatic software); now being superseded by the community-maintained OpenCCU

• MCP — Model Context Protocol specification

• ccu-ai-mcp by Mathias (mdzio) — a kindred MCP server for HomeMatic, taking a deliberately different, elegant approach (a lean Go core with user-defined HM-Script tools). See his write-up on the HomeMatic forum.

License

MIT