Deploying OpenClaw: Local Box, Home Server, Or VPS

OpenClaw is a long-running runtime, not a script you invoke and forget. It loads souls (long-lived agent personas), schedules heartbeats (periodic ticks where each soul wakes up and considers what to do), and exposes skills (capabilities it can call). All three need a home that stays on. The realistic options are your laptop, a home server, and a small VPS — each with different trade-offs.

The Docker option

OpenClaw ships a reference Dockerfile and a docker-compose.yml. Container-first deployment is the right default for any host that's not your dev laptop — it gives you reproducible installs, easy upgrades, and clean rollbacks when a release breaks something. The compose file declares the OpenClaw process, an Ollama sidecar (optional), and the volumes that hold the soul state.

# docker-compose.yml — minimal home-server deployment
services:
  openclaw:
    image: ghcr.io/openclaw/openclaw:latest
    restart: unless-stopped
    ports:
      - "8081:8081"   # mission-control web UI, bound to LAN only
    volumes:
      - ./data/souls:/var/openclaw/souls       # soul state: long-term memory
      - ./data/audit:/var/openclaw/audit       # audit log: every action, every tick
      - ./config:/etc/openclaw:ro              # config: read-only mount
    environment:
      - OPENCLAW_BACKEND=ollama
      - OPENCLAW_OLLAMA_URL=http://ollama:11434
      - OPENCLAW_LOG_LEVEL=info
    depends_on:
      - ollama

  ollama:
    image: ollama/ollama:latest
    restart: unless-stopped
    volumes:
      - ./data/ollama:/root/.ollama            # downloaded models (big)
    # No ports exposed — only OpenClaw talks to it.

Persistent volume layout

OpenClaw splits state into four directories. Mount them outside the container so an `image pull` and restart never erases anything important. The split exists for a reason — different directories have different backup and rotation policies.

• /var/openclaw/souls — durable: each soul's long-term memory, scheduled heartbeats, skill bindings. Back this up. Lose this and your souls forget who they are.
• /var/openclaw/audit — append-only: every heartbeat, every skill call, every approval gate decision. Rotate monthly; do not delete the current month.
• /var/openclaw/cache — disposable: tool-call results, model responses, anything regeneratable. Safe to wipe between releases.
• /etc/openclaw — config: soul definitions, skill registry, secrets references (NOT secrets themselves). Version-control this directory.

Restart-resilience for souls

A soul is a long-running agent — and long-running things crash. OpenClaw's design assumes restarts are routine, not exceptional. Souls write checkpoint state after each heartbeat; on boot, the runtime replays from the last checkpoint and the next scheduled tick fires on time. Your job is to make sure the host actually restarts the OpenClaw process when it falls over.

1. 1Use `restart: unless-stopped` (Docker) or `Restart=always` (systemd) so a crash auto-recovers.
2. 2Set a heartbeat watchdog: if a soul hasn't ticked in 2x its interval, log a warning. Lesson 2 covers this.
3. 3Test it deliberately — once a quarter, kill -9 the container and time how long until souls resume.
4. 4Keep migrations idempotent: a soul on version N reading a checkpoint from version N-1 must not corrupt itself.
5. 5On VPS, set the host to auto-reboot on kernel update (e.g. `unattended-upgrades` on Debian) and confirm souls come back.

The host-trust question

OpenClaw can read whatever you let it read. If it lives on a VPS, the provider can technically read your soul memory and audit logs. For a hobby weather-summary soul that's fine. For a soul that has access to your email or your bank exports, it's not. Match the data sensitivity to the host trust level — a calendar-summary soul on a $5 VPS is reasonable; an inbox-triage soul there is reckless. Lesson 3 has the full security model.

The big idea: OpenClaw's home shapes everything that follows. Pick your host on data sensitivity and uptime needs, mount volumes outside the container, and prove restart-resilience before you trust a soul with anything that matters.