OpenShell/deploy/docker/docker-compose.yml

# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# OpenShell gateway — docker-compose setup (Docker compute driver)
#
# Prerequisites:
#   - Docker Desktop (Windows / macOS) or Docker Engine + Compose plugin (Linux)
#   - The openshell CLI installed on your workstation
#
# Quick start:
#
#   1. Start the gateway:
#        docker compose up -d
#
#   2. Register the gateway with the CLI (one-time):
#        openshell gateway add http://localhost:8080 --name openshell-docker
#
#   3. Configure an AI provider (example: Anthropic):
#        ANTHROPIC_API_KEY=sk-ant-... \
#          openshell provider create --type anthropic --from-existing
#
#   4. Create a sandboxed agent — Claude Code:
#        openshell sandbox create -- claude
#      For OpenClaw, use NemoClaw: https://github.com/NVIDIA/NemoClaw
#
# Sandbox containers are managed by the gateway, not by this Compose file.
# Each `openshell sandbox create` call launches a fresh container; the gateway
# tracks their lifecycle.
#
# Configuration:
#   All gateway and driver settings live in gateway.toml in this directory.
#   Three values cannot be expressed in the TOML file and remain as env vars:
#     - OPENSHELL_DB_URL  (explicitly blocked from the config file to prevent
#                          secrets from being committed to VCS)
#     - XDG_DATA_HOME / HOME  (OS-level path-resolution vars outside the
#                               gateway config schema)
#
# command: [] note:
#   The gateway image's default CMD is ["--bind-address", "0.0.0.0", "--port",
#   "8080"].  CLI flags beat TOML in the merge order, so without clearing the
#   CMD the TOML's bind_address = "127.0.0.1:8080" is silently ignored and the
#   gateway binds 0.0.0.0.  Setting command: [] lets the TOML file own all
#   gateway settings.
#
# Data directory note:
#   /var/lib/openshell is bind-mounted at the SAME absolute path in both the
#   host and the container.  This is required so that the supervisor binary
#   extracted from the supervisor image can be passed to Docker as a host-side
#   bind-mount source when sandbox containers are created.  Named volumes
#   cannot be used here because Docker resolves bind-mount sources against the
#   host filesystem, not the container filesystem.
#
# Linux note:
#   host.docker.internal and host.openshell.internal are not automatically
#   added on Linux Docker.  Add the following under the gateway service:
#     extra_hosts:
#       - "host.docker.internal:host-gateway"
#       - "host.openshell.internal:host-gateway"

services:
  gateway:
    image: ghcr.io/nvidia/openshell/gateway:${IMAGE_TAG:-latest}
    restart: unless-stopped

    # Clear the default CMD so gateway.toml owns all settings (see note above).
    command: []

    # This setup is Docker-outside-of-Docker (DooD), not Docker-in-Docker (DinD).
    # The gateway uses the host's Docker socket to create sibling containers on the
    # host, rather than running a nested Docker daemon.  DooD does NOT require
    # --privileged; it only needs read/write access to /var/run/docker.sock.
    #
    # Run as UID 0 so the gateway can:
    #   - write the extracted supervisor binary to /var/lib/openshell
    #   - access /var/run/docker.sock (typically owned by root or the docker group)
    # Distroless images have no /etc/passwd, so the numeric UID must be used.
    # This is appropriate for local development.  Production deployments
    # should use a dedicated non-root UID with explicit docker-group membership.
    user: "0"

    ports:
      # gRPC / control-plane API. The Docker supervisor uses host networking
      # and reaches this primary loopback endpoint through the published port.
      - "127.0.0.1:${OPENSHELL_PORT:-8080}:8080"
      # Health endpoint  (GET /healthz, GET /readyz)
      - "127.0.0.1:${OPENSHELL_HEALTH_PORT:-8081}:8081"

    volumes:
      # Docker socket — lets the gateway create and manage sandbox containers.
      - /var/run/docker.sock:/var/run/docker.sock

      # Data directory — must be a bind-mount with source == target so that
      # paths written inside the container are resolvable by Docker when it
      # creates sandbox containers (see note above).
      # /var/lib/openshell is intentionally not namespaced to a sub-path
      # (e.g. /var/lib/openshell/gateway): the path must match exactly on
      # both the host and inside the container, and a single gateway per host
      # is the expected topology.
      - type: bind
        source: /var/lib/openshell
        target: /var/lib/openshell
        bind:
          create_host_path: true

      # TOML config — all gateway and driver settings live here.
      # Docker Compose resolves ./ relative to the directory that contains this
      # docker-compose.yml file (deploy/docker/), not the CWD of the caller.
      - type: bind
        source: ./gateway.toml
        target: /etc/openshell/gateway.toml
        read_only: true

    environment:
      # Point the gateway at the TOML config file mounted above.
      OPENSHELL_GATEWAY_CONFIG: /etc/openshell/gateway.toml

      # Database URL cannot be set in the TOML config file — it is explicitly
      # blocked there to prevent credential-bearing URLs (e.g. postgresql://user:pass@host/db)
      # from being committed to VCS.  A plain SQLite path contains no credentials, so it
      # is safe to include here.  Swap this for a real DSN via an env file or secret if
      # you switch to an external database.
      OPENSHELL_DB_URL: "sqlite:/var/lib/openshell/gateway.db?mode=rwc"

      # XDG path variables are OS-level; they are not part of the gateway config
      # schema.  Setting them ensures the extracted supervisor binary lands in the
      # bind-mounted directory so its path is resolvable by the host Docker daemon.
      XDG_DATA_HOME: /var/lib/openshell
      HOME: /var/lib/openshell