Deployment¶

Deployment Options¶

home-assistant-agent exposes its MCP server (console script home-assistant-mcp) four ways — local stdio, a loopback-only development listener, a least-privilege container, and a remote authenticated HTTPS boundary. Pick the row that matches where the server runs relative to your MCP client, then copy the matching mcp_config.json below. Provider endpoint, credential, selector, identity, and trust material are supplied at runtime through AgentConfig / environment variables — replace the <your-…> placeholders and never commit real values into client JSON or documentation.

#	Option	Transport	Where it runs	`mcp_config.json` key
1	stdio	`stdio`	client launches a subprocess	`command`
2	Loopback development listener	`streamable-http`	a local network port (127.0.0.1 only)	`command` or `url`
3	Least-privilege container	`stdio` or `streamable-http`	Docker / Podman on this host	`command` or `url`
4	Remote authenticated HTTPS endpoint	`streamable-http`	a remote host behind a TLS-terminating reverse proxy	`url`

1. Installed stdio process¶

The client launches the server over stdio via uvx — best for local IDEs (Cursor, Claude Desktop, VS Code):

{
  "mcpServers": {
    "home-assistant-mcp": {
      "command": "uvx",
      "args": ["--from", "home-assistant-agent", "home-assistant-mcp"],
      "env": {
        "MCP_TOOL_MODE": "intent",
        "HOME_ASSISTANT_URL": "<your-home_assistant_url>",
        "HOME_ASSISTANT_TOKEN": "<your-home_assistant_token>"
      }
    }
  }
}

2. Loopback development listener¶

Run the server as a long-lived HTTP process, bound to loopback only:

home-assistant-mcp --transport streamable-http --host 127.0.0.1 --port 8000
curl -s http://localhost:8000/health        # {"status":"OK"}

Do not expose this listener beyond loopback. Network deployments require direct TLS or an explicitly trusted TLS-terminating ingress, configured authentication, an exact MCP_ALLOWED_HOSTS, and an exact trusted-proxy CIDR policy — see Option 4.

Either let the client launch it, or connect to the already-running process by URL:

{
  "mcpServers": {
    "home-assistant-mcp": { "url": "http://localhost:8000/mcp" }
  }
}

…or, from a local checkout, run it with uv directly:

uv run home-assistant-mcp --transport streamable-http --port 8000

3. Least-privilege local container¶

Swap docker for podman for a daemonless runtime. Pin the image by digest — the shipped docker/mcp.compose.yml requires one (HOME_ASSISTANT_AGENT_MCP_IMAGE) rather than trusting a floating tag.

(a) Launch directly from mcp_config.json (stdio over the container — no ports to manage):

{
  "mcpServers": {
    "home-assistant-mcp": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "--read-only", "--cap-drop=ALL", "--security-opt=no-new-privileges",
        "--pids-limit=256", "--tmpfs", "/tmp:rw,noexec,nosuid,nodev,size=64m",
        "-e", "TRANSPORT=stdio",
        "-e", "HOME_ASSISTANT_URL=<your-home_assistant_url>",
        "-e", "HOME_ASSISTANT_TOKEN=<your-home_assistant_token>",
        "registry.example.invalid/home-assistant-agent@sha256:<digest>",
        "home-assistant-mcp"
      ]
    }
  }
}

The operator projects the selected AgentConfig profile into the process at runtime; the image remains immutable and contains no environment connection profile.

(b) Run a long-lived streamable-http container, then connect by URL:

docker compose -f docker/mcp.compose.yml up -d   # from a clone of this repo

{
  "mcpServers": {
    "home-assistant-mcp": { "url": "http://localhost:8000/mcp" }
  }
}

4. Remote authenticated HTTPS endpoint¶

When the server is deployed remotely and published through a TLS-terminating reverse proxy, connect with the "url" key — no local process or image required:

{
  "mcpServers": {
    "home-assistant": {"url": "https://service.example.invalid/mcp"}
  }
}

Store the real remote URL, outbound identity reference, and TLS-profile reference in AgentConfig, not in MCP client JSON or documentation. See Behind a Caddy reverse proxy and DNS with Technitium below for standing one up.

This page covers running home-assistant-agent as a long-lived server: the transports, a Docker Compose stack, the optional A2A agent server, putting it behind a Caddy reverse proxy, and giving it a DNS name with Technitium. To provision the Home Assistant instance it connects to, see Backing Platform.

home-assistant-agent ships two entry points: an MCP server (console script home-assistant-mcp) — a typed, deterministic tool surface a policy router or agent calls — and an optional A2A agent server (console script home-assistant-agent) that wraps that tool surface in a Pydantic-AI graph agent.

Run the MCP server¶

The transport is selected with --transport (or the TRANSPORT env var):

stdio (default)streamable-httpsse

home-assistant-mcp

For IDE / desktop MCP clients that launch the server as a subprocess.

home-assistant-mcp --transport streamable-http --host 0.0.0.0 --port 8000

A network server with a /health endpoint and /mcp route.

home-assistant-mcp --transport sse --host 0.0.0.0 --port 8000

Health check (HTTP transports):

curl -s http://localhost:8000/health        # {"status":"OK"}

Configuration (environment)¶

home-assistant-agent is configured entirely from the environment. The required set:

Var	Default	Meaning
`HOME_ASSISTANT_URL`	`http://localhost:8123`	Home Assistant base URL
`HOME_ASSISTANT_TOKEN`	(unset)	Long-lived access token
`TLS_PROFILE` / `TLS_PROFILE_REF`	(system trust)	AgentConfig private-CA/mTLS profile; verification remains mandatory
`HOST`	`0.0.0.0`	Bind address (HTTP transports)
`PORT`	`8000`	Bind port (HTTP transports)
`TRANSPORT`	`stdio`	`stdio`, `streamable-http`, or `sse`

Each MCP tool domain is registered only when its flag is enabled (all default to True): CONFIGTOOL, STATESTOOL, SERVICESTOOL, EVENTSTOOL, HISTORYTOOL, LOGBOOKTOOL, CALENDARTOOL, PANELSTOOL, VOICETOOL, ENTITIESTOOL, SYSTEMTOOL. The full set, with the agent and observability settings, is documented in .env.example. Copy it to .env and fill in only what you use.

Docker Compose¶

The repo ships docker/mcp.compose.yml. It reads a sibling .env and publishes the HTTP server on :8000:

services:
  home-assistant-agent-mcp:
    image: example/home-assistant-agent@sha256:<digest>
    container_name: home-assistant-agent-mcp
    hostname: home-assistant-agent-mcp
    restart: always
    env_file:
      - ../.env
    environment:
      - PYTHONUNBUFFERED=1
      - HOST=0.0.0.0
      - PORT=8000
      - TRANSPORT=streamable-http
    ports:
      - "8000:8000"
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
      interval: 30s
      timeout: 10s
      retries: 3

cp .env.example .env          # then edit HOME_ASSISTANT_* values
docker compose -f docker/mcp.compose.yml up -d
docker compose -f docker/mcp.compose.yml logs -f

Run the A2A agent server¶

For an autonomous agent over the same tool surface, run the agent server (console script home-assistant-agent). It connects to the MCP server over HTTP and exposes an A2A endpoint plus an embedded web UI.

export HOME_ASSISTANT_URL=http://your-home-assistant:8123
export HOME_ASSISTANT_TOKEN=your_long_lived_access_token
home-assistant-agent --provider openai --model-id gpt-4o --api-key sk-...

The repo ships docker/agent.compose.yml, which runs the MCP server and the agent server on one network. The agent reaches the MCP server by container name via MCP_URL and publishes the A2A server on :9004:

services:
  home-assistant-agent-mcp:
    image: example/home-assistant-agent@sha256:<digest>
    hostname: home-assistant-agent-mcp
    env_file: ["../.env"]
    environment:
      - TRANSPORT=streamable-http
      - HOST=0.0.0.0
      - PORT=8000
    ports: ["8000:8000"]

  home-assistant-agent-agent:
    image: example/home-assistant-agent@sha256:<digest>
    depends_on: [home-assistant-agent-mcp]
    command: ["home-assistant-agent"]
    env_file: ["../.env"]
    environment:
      - HOST=0.0.0.0
      - PORT=9004
      - MCP_URL=http://home-assistant-agent-mcp:8000/mcp
      - PROVIDER=${PROVIDER:-openai}
      - MODEL_ID=${MODEL_ID:-gpt-4o}
      - ENABLE_WEB_UI=True
    ports: ["9004:9004"]

docker compose -f docker/agent.compose.yml up -d

Behind a Caddy reverse proxy¶

Expose the HTTP server on a hostname with automatic TLS. Add to your Caddyfile:

# Internal (self-signed) — homelab .example.invalid zone
home-assistant-agent.example.invalid {
    tls internal
    reverse_proxy home-assistant-agent-mcp:8000
}

# Public — automatic Let's Encrypt
home-assistant-agent.example.com {
    reverse_proxy home-assistant-agent-mcp:8000
}

Reload Caddy:

docker compose -f services/caddy/compose.yml exec caddy caddy reload --config /etc/caddy/Caddyfile

DNS with Technitium¶

Point the hostname at the host running Caddy. Via the Technitium API:

curl -s "http://technitium.example.invalid:5380/api/zones/records/add" \
  --data-urlencode "token=$TECHNITIUM_DNS_TOKEN" \
  --data-urlencode "domain=home-assistant-agent.example.invalid" \
  --data-urlencode "zone=arpa" \
  --data-urlencode "type=A" \
  --data-urlencode "ipAddress=192.0.2.10" \
  --data-urlencode "ttl=3600"

…or add an A record home-assistant-agent.example.invalid → <caddy-host-ip> in the Technitium web console (http://technitium.example.invalid:5380). The ecosystem technitium-dns-mcp automates this as a tool.

Register with an MCP client¶

Add to your client's mcp_config.json:

{
  "mcpServers": {
    "home-assistant-agent": {
      "command": "uv",
      "args": ["run", "home-assistant-mcp"],
      "env": {
        "HOME_ASSISTANT_URL": "http://your-home-assistant:8123",
        "HOME_ASSISTANT_TOKEN": "your_long_lived_access_token"
      }
    }
  }
}

For a remote HTTP server, point the client at http://home-assistant-agent.example.invalid/mcp instead.