stack/docker-compose.yml

name: arr-stack

################################################################################
# Arr stack
#   gluetun      WireGuard VPN client + firewall killswitch
#   qbittorrent  torrent client, lives INSIDE gluetun's network namespace
#   prowlarr     indexer manager (torrents + NZB/Usenet indexers)
#   sabnzbd      Usenet NZB download client (Newshosting, Tweaknews, …)
#   sonarr       TV         radarr    movies       bazarr    subtitles
#   jellyfin     media server
#
# AI / coding stack (local, no VPN)
#   bonsai       Prism Ternary-Bonsai-27B via llama-server (OpenAI-compatible API)
#   open-webui   Chat UI + RAG + tools + code interpreter
#   searxng      Private metasearch for web RAG
#   forgejo      Self-hosted Git (repos for coding + OpenHands)
#   openhands    Optional autonomous coding agent (profile: coding)
#
# Lockdown model:
#   * qbittorrent uses `network_mode: service:gluetun` -> it has NO network
#     interface of its own. Every packet it sends/receives goes through
#     gluetun, and gluetun's built-in firewall only lets traffic out via the
#     WireGuard tunnel. Tunnel down = zero connectivity. No leaks possible.
#   * Do NOT publish qBittorrent's BitTorrent listening port on the host -
#     inbound peer connections must come through the tunnel (port forwarding),
#     never via your real IP.
#   * AI services stay on arr-net (LAN only). They do NOT route through gluetun.
################################################################################

networks:
  arr-net:
    driver: bridge

services:

  # ------------------------------------------------------------- VPN gateway #
  gluetun:
    image: qmcgaw/gluetun:v3
    container_name: gluetun
    restart: unless-stopped
    cap_add:
      - NET_ADMIN
    devices:
      - /dev/net/tun:/dev/net/tun
    networks:
      - arr-net
    ports:
      # Ports for anything running inside gluetun's namespace are published
      # HERE, on gluetun - not on the inner container.
      - 8080:8080/tcp            # qBittorrent WebUI
    volumes:
      # Your provider's WireGuard config, renamed to wg0.conf
      # (see wireguard/wg0.conf.example and the README for required edits)
      - ./wireguard/wg0.conf:/gluetun/wireguard/wg0.conf:ro
    environment:
      VPN_SERVICE_PROVIDER: custom
      VPN_TYPE: wireguard
      TZ: ${TZ}
      # Lets your LAN reach the qBittorrent WebUI through gluetun's firewall
      FIREWALL_OUTBOUND_SUBNETS: ${LAN_SUBNET}
      # --- optional: provider port forwarding (better swarm connectivity) ---
      # FIREWALL_VPN_INPUT_PORTS: "12345"     # port your provider assigned you
      # VPN_PORT_FORWARDING: "on"             # for providers with NAT-PMP API
      # VPN_PORT_FORWARDING_PROVIDER: protonvpn

  # --------------------------------------------------------- torrent client #
  qbittorrent:
    image: lscr.io/linuxserver/qbittorrent:latest
    container_name: qbittorrent
    restart: unless-stopped
    network_mode: service:gluetun          # <-- the entire isolation trick
    depends_on:
      gluetun:
        condition: service_healthy         # start only once the tunnel is up
    environment:
      PUID: ${PUID}
      PGID: ${PGID}
      TZ: ${TZ}
      WEBUI_PORT: "8080"
    volumes:
      - ${CONFIG_DIR}/qbittorrent:/config
      - ${DATA_DIR}:/data                  # single shared tree -> hardlinks

  # -------------------------------------------------------------- indexers #
  prowlarr:
    image: lscr.io/linuxserver/prowlarr:latest
    container_name: prowlarr
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 9696:9696
    environment: &arr-env
      PUID: ${PUID}
      PGID: ${PGID}
      TZ: ${TZ}
    volumes:
      - ${CONFIG_DIR}/prowlarr:/config
    depends_on:
      - flaresolverr

  # Cloudflare / JS challenge proxy for Prowlarr indexers (not a general captcha solver)
  # Wire in Prowlarr: Settings → Indexers → FlareSolverr → http://flaresolverr:8191
  flaresolverr:
    image: ghcr.io/flaresolverr/flaresolverr:latest
    container_name: flaresolverr
    restart: unless-stopped
    networks: [arr-net]
    environment:
      LOG_LEVEL: ${FLARESOLVERR_LOG_LEVEL:-info}
      LOG_HTML: "false"
      CAPTCHA_SOLVER: ${FLARESOLVERR_CAPTCHA_SOLVER:-none}
      TZ: ${TZ}
    # Optional LAN debug UI/API — leave commented to keep it internal-only
    # ports:
    #   - "8191:8191"

  # ---------------------------------------------------------- usenet / NZB #
  # SABnzbd talks to Usenet providers over SSL (no VPN required).
  # Incomplete + complete under /data/usenet so Sonarr/Radarr can hardlink.
  sabnzbd:
    image: lscr.io/linuxserver/sabnzbd:latest
    container_name: sabnzbd
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - "${SABNZBD_PORT:-8085}:8080"
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/sabnzbd:/config
      - ${DATA_DIR}:/data

  # -------------------------------------------------------------------- TV #
  sonarr:
    image: lscr.io/linuxserver/sonarr:latest
    container_name: sonarr
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 8989:8989
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/sonarr:/config
      - ${DATA_DIR}:/data

  # ----------------------------------------------------------------- movies #
  radarr:
    image: lscr.io/linuxserver/radarr:latest
    container_name: radarr
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 7878:7878
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/radarr:/config
      - ${DATA_DIR}:/data

  # -------------------------------------------------------------- subtitles #
  bazarr:
    image: lscr.io/linuxserver/bazarr:latest
    container_name: bazarr
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 6767:6767
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/bazarr:/config
      - ${DATA_DIR}:/data

  # ----------------------------------------------------------- media server #
  jellyfin:
    image: lscr.io/linuxserver/jellyfin:latest
    container_name: jellyfin
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 8096:8096/tcp           # web UI / clients
      # - 8920:8920/tcp         # optional HTTPS
      # - 7359:7359/udp         # optional client autodiscovery
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/jellyfin:/config
      - ${DATA_DIR}/media:/data/media:ro   # read-only: Jellyfin never touches files
    # --- optional: Intel QuickSync hardware transcoding --- (Unused on this AMD machine)
    # devices:
    #   - /dev/dri:/dev/dri

  # ------------------------------------------------------------------- music #
  lidarr:
    image: lscr.io/linuxserver/lidarr:latest
    container_name: lidarr
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - 8686:8686
    environment: *arr-env
    volumes:
      - ${CONFIG_DIR}/lidarr:/config
      - ${DATA_DIR}:/data                  # same /data tree as Sonarr/Radarr

  # ==========================================================================
  # Optional media extras - uncomment as needed
  # ==========================================================================
  #
  # jellyseerr:                              # request UI for Jellyfin
  #   image: fallenbagel/jellyseerr:latest
  #   container_name: jellyseerr
  #   restart: unless-stopped
  #   networks: [arr-net]
  #   ports: ["5055:5055"]
  #   environment: *arr-env
  #   volumes:
  #     - ${CONFIG_DIR}/jellyseerr:/app/config

  # ==========================================================================
  # Local AI — Ternary-Bonsai-27B + Open WebUI + RAG + coding helpers
  # Target: Beelink SER5 PRO (Ryzen 7 7735HS + Radeon 680M, no NVIDIA)
  # ==========================================================================

  # OpenAI-compatible API for Ternary-Bonsai-27B (PrismML Q2_0 GGUF)
  # First:  ./scripts/download-bonsai-model.sh
  # Then:   docker compose up -d --build bonsai open-webui searxng
  bonsai:
    build:
      context: ./ai
      dockerfile: Dockerfile
      args:
        # cpu = reliable default on SER5 PRO
        # vulkan = try Radeon 680M offload (also set NGL=99 and uncomment devices)
        BONSAI_BACKEND: ${BONSAI_BACKEND:-vulkan}
    image: arr-stack/bonsai:local
    container_name: bonsai
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - "${BONSAI_PORT:-8081}:8080"   # avoid clash with qBittorrent :8080
    environment:
      TZ: ${TZ}
      MODEL_PATH: ${BONSAI_MODEL_PATH:-/models/Ternary-Bonsai-27B-Q2_0.gguf}
      # 0 = CPU (recommended start). Vulkan image: try 99 with /dev/dri mounted.
      NGL: ${BONSAI_NGL:-99}
      # 16k is a good default on 32–64 GB systems with media stack co-resident.
      # Long docs / full-repo: try 32768 or 65536; enable KV4=1 if RAM is tight.
      CTX_SIZE: ${BONSAI_CTX:-16384}
      TEMP: ${BONSAI_TEMP:-0.7}
      TOP_P: ${BONSAI_TOP_P:-0.95}
      TOP_K: ${BONSAI_TOP_K:-20}
      # 0 = auto; or set to physical cores (e.g. 8 on 7735HS) leaving some for Arrs
      THREADS: ${BONSAI_THREADS:-0}
      KV4: ${BONSAI_KV4:-0}
      EXTRA_ARGS: ${BONSAI_EXTRA_ARGS:-}
    volumes:
      - ${BONSAI_MODELS_DIR:-./models/bonsai}:/models:ro
    # Shared memory helps large context / mmap
    shm_size: "4gb"
    # Vulkan / AMD iGPU: devices + group_add live ONLY in
    # docker-compose.override.yml (Compose merges lists and rejects duplicates).
    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://127.0.0.1:8080/health"]
      interval: 30s
      timeout: 10s
      start_period: 180s
      retries: 5

  # ChatGPT-style UI with built-in RAG, tools, and code interpreter
  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    container_name: open-webui
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - "${OPEN_WEBUI_PORT:-3000}:8080"
    depends_on:
      bonsai:
        condition: service_healthy
      searxng:
        condition: service_started
    environment:
      TZ: ${TZ}
      # Point at local llama-server (OpenAI-compatible)
      OPENAI_API_BASE_URL: http://bonsai:8080/v1
      OPENAI_API_KEY: ${OPENAI_API_KEY:-sk-local-bonsai}
      # Disable Ollama unless you add it later
      ENABLE_OLLAMA_API: "false"
      # First account becomes admin; set WEBUI_AUTH=false only on trusted LAN
      WEBUI_AUTH: ${WEBUI_AUTH:-true}
      # RAG: local embeddings run inside open-webui (downloads on first use)
      RAG_EMBEDDING_ENGINE: ${RAG_EMBEDDING_ENGINE:-}
      CHUNK_SIZE: ${RAG_CHUNK_SIZE:-1000}
      CHUNK_OVERLAP: ${RAG_CHUNK_OVERLAP:-100}
      # Web search via local SearXNG
      ENABLE_RAG_WEB_SEARCH: "true"
      ENABLE_SEARCH_QUERY: "true"
      RAG_WEB_SEARCH_ENGINE: searxng
      SEARXNG_QUERY_URL: http://searxng:8080/search?q=<query>
      # Allow fetching internal / LAN URLs for RAG (homelab docs)
      ENABLE_RAG_LOCAL_WEB_FETCH: "true"
      # Code interpreter (runs in open-webui sandbox)
      ENABLE_CODE_INTERPRETER: "true"
      # Reduce background LLM spam on a single local model
      ENABLE_TITLE_GENERATION: "false"
      ENABLE_FOLLOW_UP_GENERATION: "false"
      ENABLE_TAGS_GENERATION: "false"
      # Default model id as reported by llama-server /v1/models
      DEFAULT_MODELS: ${OPENWEBUI_DEFAULT_MODEL:-}
    volumes:
      - ${CONFIG_DIR}/open-webui:/app/backend/data
      # Your local files for RAG / "read my project" workflows
      - ${AI_WORKSPACE_DIR:-./workspace}:/workspace:ro
      # Optional: expose media library read-only for Q&A about your library tree
      # - ${DATA_DIR}/media:/media:ro
    extra_hosts:
      - "host.docker.internal:host-gateway"

  # Private metasearch (no API keys) for Open WebUI web RAG
  searxng:
    image: docker.io/searxng/searxng:latest
    container_name: searxng
    restart: unless-stopped
    networks: [arr-net]
    # Bind to localhost only — Open WebUI reaches it on arr-net
    ports:
      - "127.0.0.1:${SEARXNG_PORT:-8888}:8080"
    environment:
      SEARXNG_BASE_URL: http://localhost:${SEARXNG_PORT:-8888}/
      SEARXNG_SECRET: ${SEARXNG_SECRET:-change-me-to-a-long-random-string}
    volumes:
      - ./searxng:/etc/searxng:rw
    cap_drop:
      - ALL
    cap_add:
      - CHOWN
      - SETGID
      - SETUID

  # -------------------------------------------------------------------------
  # Forgejo — self-hosted Git (Gitea soft-fork). SQLite is fine for a homelab.
  # Web: http://host:3002   Git SSH: host:2222  (git@host:user/repo.git with port 2222)
  # First visit: create admin account, then create repos / push code.
  # -------------------------------------------------------------------------
  forgejo:
    image: codeberg.org/forgejo/forgejo:11
    container_name: forgejo
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - "${FORGEJO_HTTP_PORT:-3002}:3000"
      - "${FORGEJO_SSH_PORT:-2222}:22"
    environment:
      USER_UID: ${PUID:-1000}
      USER_GID: ${PGID:-1000}
      TZ: ${TZ}
      # Image already runs OpenSSH on :22 (mapped to host FORGEJO_SSH_PORT).
      # Do NOT enable Forgejo's built-in SSH too — both bind :22 and the
      # process fatals with "address already in use" (connection refused on :3002).
      FORGEJO__server__START_SSH_SERVER: "false"
      FORGEJO__server__DISABLE_SSH: "false"
      FORGEJO__server__SSH_PORT: ${FORGEJO_SSH_PORT:-2222}
      FORGEJO__server__SSH_DOMAIN: ${FORGEJO_DOMAIN:-192.168.8.123}
      FORGEJO__server__DOMAIN: ${FORGEJO_DOMAIN:-192.168.8.123}
      FORGEJO__server__ROOT_URL: ${FORGEJO_ROOT_URL:-http://192.168.8.123:3002/}
      FORGEJO__server__HTTP_PORT: "3000"
      # SQLite by default (no extra DB container)
      FORGEJO__database__DB_TYPE: sqlite3
      FORGEJO__service__DISABLE_REGISTRATION: ${FORGEJO_DISABLE_REGISTRATION:-false}
      FORGEJO__actions__ENABLED: "true"
    volumes:
      - ${CONFIG_DIR}/forgejo:/data
      - /etc/localtime:/etc/localtime:ro

  # -------------------------------------------------------------------------
  # OpenHands — autonomous coding agent (opt-in)
  # Start with:  docker compose --profile coding up -d openhands
  # UI: http://host:3001  → Settings → custom model openai/<id>
  #     Base URL: http://bonsai:8080/v1   API key: sk-local-bonsai
  # Needs Docker socket to spawn sandboxed runtimes.
  # -------------------------------------------------------------------------
  openhands:
    profiles: ["coding"]
    # Prefer GHCR — docker.all-hands.dev often fails DNS / is decommissioned
    image: ghcr.io/all-hands-ai/openhands:0.54
    container_name: openhands
    restart: unless-stopped
    networks: [arr-net]
    ports:
      - "${OPENHANDS_PORT:-3001}:3000"
    extra_hosts:
      - "host.docker.internal:host-gateway"
    environment:
      TZ: ${TZ}
      SANDBOX_RUNTIME_CONTAINER_IMAGE: ghcr.io/all-hands-ai/runtime:0.54-nikolaik
      LOG_ALL_EVENTS: "true"
      # Pre-wire local Bonsai (override in UI if model id differs)
      LLM_BASE_URL: http://bonsai:8080/v1
      LLM_API_KEY: ${OPENAI_API_KEY:-sk-local-bonsai}
      LLM_MODEL: ${OPENHANDS_MODEL:-openai/local}
      WORKSPACE_MOUNT_PATH: ${AI_WORKSPACE_DIR:-./workspace}
      SANDBOX_VOLUMES: ${AI_WORKSPACE_DIR:-./workspace}:/workspace:rw
    volumes:
      - /var/run/docker.sock:/var/run/docker.sock
      - ${CONFIG_DIR}/openhands:/.openhands-state
      - ${AI_WORKSPACE_DIR:-./workspace}:/opt/workspace_base
    depends_on:
      bonsai:
        condition: service_healthy