Deployment

This guide covers deploying OpenClaw on your own infrastructure, running it as a system daemon, and planning for growth.

Recommended Topology

The simplest production deployment is a single server running the OpenClaw Gateway with all agents:

Your Server (Ubuntu / Linux)
├── OpenClaw Gateway (:18789)
│   ├── coordinator agent
│   ├── dev-agent (Docker sandboxed)
│   ├── product-agent
│   └── risk-agent
├── Tailscale / SSH tunnel → remote access
├── Database (read-only replica) → for agent queries
└── systemd service → auto-restart, logging

Installation

Install OpenClaw as a systemd daemon:

openclaw onboard --install-daemon

This sets up:

• Auto-restart on failure
• System logging via journald
• Automatic startup on boot

Starting the Gateway

openclaw gateway --port 18789

Or, if installed as a daemon:

systemctl start openclaw
systemctl status openclaw

Remote Access

If your server is in a local datacenter or behind a firewall, use Tailscale Serve or an SSH tunnel with token auth for remote access.

Health Checks

# Full system health check
openclaw doctor

# Check all channel connections
openclaw channels status --probe

# Verify agent routing
openclaw agents list --bindings

Scaling Considerations

When to Stay on a Single Server

A single server handles most use cases well:

• Up to 5–10 agents
• Moderate message volume
• Team of 10–50 users

When to Consider Scaling

• Token costs exceeding your budget on the current setup
• Uptime requirements beyond what a single server provides
• Need for geographic distribution

Scaling Options

1. Dedicated VPS — Some cloud providers offer one-click OpenClaw deploys for straightforward scaling
2. Load-balanced gateways — Multiple gateway instances behind a load balancer (advanced)
3. Separate gateways per team — If teams are fully independent, separate installations may be simpler

Hardware Recommendations

Dedicated machine required. OpenClaw has full filesystem access and can execute arbitrary commands. Never install it on a machine with sensitive data or active personal use.

• Mac Mini (M4, 16GB, 256GB) — Quiet, low-power, reliable. Good for home/office setups.
• VPS — DigitalOcean, Railway, Render, or Google Cloud. Best for teams that need uptime without managing hardware.
• Old laptop — Fine for experimentation. Keep it plugged in and connected to the network.

For physical hardware without a monitor: enable Screen Sharing and Remote Login in macOS System Settings, then SSH in from your main machine.

Cost Considerations

Token costs depend on model choices, heartbeat frequency, and message volume.

• 1–2 agents on Sonnet — ~$100–200/month via API
• 5–10 agents (mixed Opus + Sonnet) — ~$500–1000/month via API
• Subscription-based — Using an existing Claude or ChatGPT subscription is cheapest (~$20/month) but check the provider's terms of service for automated usage

Cost levers:

• Model-per-agent — Use Opus only for agents that need complex reasoning (coordinator, risk). Use Sonnet for everything else.
• Sub-agents on cheaper models — The coordinator does synthesis; sub-agents do data gathering on Sonnet.
• Heartbeat frequency — Fewer checks means fewer tokens. Most agents don't need checks more frequent than every 30 minutes.
• lightContext flag — Set lightContext: true on heartbeat tasks that don't need full conversation history.

Troubleshooting

Agent not responding:

1. openclaw doctor — Check gateway health
2. openclaw channels status --probe — Verify channel connections
3. openclaw agents list --bindings — Confirm the agent is bound to the right channel

Agent forgetting tasks:

• Check that the session memory hook is enabled
• Ask the agent to inspect its own crons and HEARTBEAT.md
• Review MEMORY.md for unexpected or missing entries

Agent executing wrong actions:

• Review TOOLS.md for incorrect tool references
• Check SOUL.md boundaries — they may need to be more explicit
• Run openclaw security audit to check for permission gaps

Broken scheduling:

• Ask the agent to inspect and list its cron definitions
• Verify HEARTBEAT.md has the correct format and intervals
• Check that the heartbeat interval in openclaw.json is set (default: 30 minutes)

Recovery from severe breakage: OpenClaw workspaces are just markdown files and JSON configs. You can debug them manually or use Claude Code to diagnose issues by pasting the workspace contents.

Phased Rollout

Phase 1: Foundation (Weeks 1–2)

1. Install OpenClaw on your server
2. Set up a single coordinator agent
3. Connect to one messaging platform
4. Write SOUL.md, AGENTS.md, USER.md
5. Build a simple periodic workflow with HEARTBEAT.md
6. Validate memory, sessions, and heartbeat behavior

Phase 2: Second Agent (Weeks 3–4)

1. Add a specialist agent (e.g., engineering)
2. Create its workspace with role-specific persona files
3. Bind to the appropriate channel
4. Enable Docker sandbox for code-executing agents
5. Test workflows specific to that role

Phase 3: Full Team (Weeks 5–6)

1. Add remaining specialist agents
2. Create their workspaces and channel bindings
3. Build custom skills (SQL runner, documentation writer, etc.)
4. Test cross-agent data sharing via files

Phase 4: Orchestration (Weeks 7–8)

1. Enable sessions_spawn for the coordinator agent
2. Build sub-agent workflows (weekly brief, analysis, monitoring)
3. Set up cron jobs for automated reports
4. Add additional agents if the team is ready

Phase 5: Ongoing Optimization

1. Monitor token costs per agent
2. Tune HEARTBEAT.md intervals
3. Iterate on workspace files based on real usage
4. Expand skills from the ClawHub registry as needed