homeclaw/docker-compose.yml

version: "3.9"

# ============================================================
# SmallClaw / LocalClaw – docker-compose.yml
#
# Supported providers (set SMALLCLAW_PROVIDER in .env):
#   ollama        – bundled Ollama container (default)
#   lm_studio     – LM Studio on your HOST machine (port 1234)
#   llama_cpp     – llama.cpp server on your HOST machine (port 8080)
#   openai        – OpenAI API key (cloud)
#   openai_codex  – OpenAI OAuth / ChatGPT Plus (cloud)
#
# Quick start:
#   cp .env.example .env        # then edit .env for your provider
#   docker compose up -d        # start everything
#   docker compose logs -f      # follow logs
#   docker compose down         # stop & remove containers
#   docker compose down -v      # also wipe volumes (full reset)
# ============================================================

services:

  # ── Ollama ────────────────────────────────────────────────
  # Only relevant when SMALLCLAW_PROVIDER=ollama.
  # If you're using lm_studio / llama_cpp / openai / openai_codex
  # you can comment out or remove the ollama + model-init services.
  ollama:
    image: ollama/ollama:latest
    container_name: smallclaw-ollama
    restart: unless-stopped
    profiles:
      - ollama   # start only when using: docker compose --profile ollama up
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama
    environment:
      - OLLAMA_HOST=0.0.0.0
    # ── GPU support ──────────────────────────────────────────
    # NVIDIA (requires nvidia-container-toolkit on the host):
    # deploy:
    #   resources:
    #     reservations:
    #       devices:
    #         - driver: nvidia
    #           count: all
    #           capabilities: [gpu]
    #
    # AMD / ROCm:
    # devices:
    #   - /dev/kfd
    #   - /dev/dri
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:11434/api/tags"]
      interval: 20s
      timeout: 10s
      retries: 5
      start_period: 10s

  # ── Model pull (one-shot init, Ollama only) ──────────────
  model-init:
    image: ollama/ollama:latest
    container_name: smallclaw-model-init
    profiles:
      - ollama
    depends_on:
      ollama:
        condition: service_healthy
    volumes:
      - ollama_data:/root/.ollama
    environment:
      - OLLAMA_HOST=http://ollama:11434
      - DEFAULT_MODEL=${SMALLCLAW_DEFAULT_MODEL:-qwen3:4b}
    entrypoint: >
      sh -c "
        echo '>>> Pulling model: '$$DEFAULT_MODEL;
        ollama pull $$DEFAULT_MODEL;
        echo '>>> Done.';
      "
    restart: "no"

  # ── SmallClaw Gateway ────────────────────────────────────
  smallclaw:
    build:
      context: .
      dockerfile: Dockerfile
    container_name: smallclaw-app
    restart: unless-stopped
    ports:
      - "${HOST_PORT:-18789}:18789"

    # Allow the container to reach LM Studio / llama.cpp on the HOST.
    # On Linux, host.docker.internal isn't automatically available so we
    # inject it via extra_hosts. On Mac/Windows Docker Desktop it works
    # out of the box, but adding it here doesn't hurt.
    extra_hosts:
      - "host.docker.internal:host-gateway"

    volumes:
      - smallclaw_data:/data
      - smallclaw_workspace:/data/workspace
      # OpenAI Codex OAuth tokens are stored in ~/.localclaw on your host.
      # Mount the directory so tokens survive container restarts and the
      # initial `smallclaw auth login` can be run once on the host.
      # Comment this out if you're not using openai_codex.
      - ${LOCALCLAW_CONFIG_DIR:-~/.localclaw}:/root/.localclaw

    environment:
      - NODE_ENV=production
      - DOCKER_CONTAINER=true
      - GATEWAY_PORT=18789
      - GATEWAY_HOST=0.0.0.0
      - SMALLCLAW_DATA_DIR=/data
      - SMALLCLAW_WORKSPACE_DIR=/data/workspace

      # ── Active provider ──────────────────────────────────
      - SMALLCLAW_PROVIDER=${SMALLCLAW_PROVIDER:-ollama}

      # ── Ollama ───────────────────────────────────────────
      # Points to the bundled container by default.
      # Override in .env: OLLAMA_HOST=http://host.docker.internal:11434
      # to use Ollama running on your host machine instead.
      - OLLAMA_HOST=${OLLAMA_HOST:-http://ollama:11434}

      # ── LM Studio ────────────────────────────────────────
      # Reaches LM Studio running on the host via host.docker.internal.
      - LM_STUDIO_ENDPOINT=${LM_STUDIO_ENDPOINT:-http://host.docker.internal:1234}
      - LM_STUDIO_API_KEY=${LM_STUDIO_API_KEY:-}
      - LM_STUDIO_MODEL=${LM_STUDIO_MODEL:-}

      # ── llama.cpp ────────────────────────────────────────
      - LLAMA_CPP_ENDPOINT=${LLAMA_CPP_ENDPOINT:-http://host.docker.internal:8080}
      - LLAMA_CPP_MODEL=${LLAMA_CPP_MODEL:-}

      # ── OpenAI ───────────────────────────────────────────
      - OPENAI_API_KEY=${OPENAI_API_KEY:-}
      - OPENAI_MODEL=${OPENAI_MODEL:-gpt-4o}

      # ── OpenAI Codex (OAuth) ─────────────────────────────
      # Tokens live in the mounted ~/.localclaw volume above.
      - CODEX_MODEL=${CODEX_MODEL:-gpt-5.3-codex}

    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:18789/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 20s

# ── Named volumes ────────────────────────────────────────────
volumes:
  ollama_data:
    driver: local
  smallclaw_data:
    driver: local
  smallclaw_workspace:
    driver: local