Master OpenClaw: Install, Configure, and Scale Multi‑Agent AI Automation

What is OpenClaw?

OpenClaw is an AI‑agent platform that orchestrates external large‑language‑model (LLM) providers. It does not contain a model itself; instead it schedules calls to configured LLM APIs and provides tools for file I/O, command execution, browser automation, and scheduled tasks.

Core Components

Gateway – central message hub that listens on port 18789 by default.

Channel – entry/exit point for messages (e.g., Feishu, Telegram, QQ).

Agent – a worker with its own identity (SOUL.md), behavior (AGENTS.md), memory (MEMORY.md) and a dedicated workspace.

Tool – built‑in utilities such as file read/write, shell execution, browser control, and cron scheduling.

Skill – optional plug‑ins that extend an agent’s capabilities.

Configuration Files

All configuration lives under ~/.openclaw/:

~/.openclaw/
  openclaw.json      # main JSON configuration (gateway, agents, tools, skills)
  skills/            # global skill directory
  agents/            # per‑agent state and sessions
  secrets/           # encrypted API keys
  workspace/         # default workspace for agents

Installation

Install Node.js 22 or newer. On macOS you can use Homebrew: brew install node On Windows download the installer from https://nodejs.org/.

Run the platform installer:

# macOS
curl -fsSL https://openclaw.ai/install.sh | bash
# Windows PowerShell
iwr -useb https://openclaw.ai/install.ps1 | iex

Run the onboarding wizard to select a model, store API keys, and bind a channel:

openclaw onboard --install-daemon

Token‑Cost Management

Token pricing varies by model (e.g., qwen‑plus‑latest costs $0.8 / M input tokens and $2 / M output tokens). The system prompt consumes roughly 13 k tokens (~43 k English characters). Reduce cost by:

Setting reserveTokensFloor and ttl in openclaw.json to prune context.

Keeping SOUL.md and AGENTS.md concise.

Using local LLMs via Ollama to avoid API fees.

Splitting long conversations into new sessions with /new or /reset.

Basic Usage Scenarios

Web UI: openclaw dashboard Terminal UI (TUI): openclaw tui Channel commands after binding credentials (Feishu, Telegram, etc.).

Cron job example:

openclaw cron add \
  --name "MorningGreeting" \
  --cron "0 8 * * *" \
  --message "Good morning! This is your AI assistant." \
  --channel feishu \
  --announce

Engineering Practices (Multi‑Agent)

Create isolated agents with separate workspaces for complex workflows:

openclaw agents add leader --workspace "~/.openclaw/workspace-leader"
openclaw agents add worker --workspace "~/.openclaw/workspace-worker"

Assign different models to each agent in openclaw.json and bind each to a distinct channel. This prevents cross‑talk and keeps resources scoped.

Security Recommendations

Never store API keys in plain text; use the web UI to encrypt them.

Set dmPolicy to pairing or allowlist to block unsolicited private messages.

Restrict tool usage, e.g. deny browser and exec tools globally:

{
  "tools": {
    "deny": ["browser", "exec"]
  }
}

Limit file‑system access to the workspace only:

{
  "tools": {
    "exec": {
      "applyPatch": {"workspaceOnly": true}
    }
  }
}

Audit the configuration:

openclaw security audit --deep

Common Troubleshooting

Gateway fails to start – force‑release the port and restart:

openclaw gateway --force
openclaw gateway restart

Model connection errors – test the API and reload secrets:

openclaw api test
openclaw secrets reload

Configuration changes not applied – restart the gateway if hot‑reload does not pick up the change:

openclaw gateway restart

Reference Resources

Official documentation: https://docs.openclaw.ai

GitHub repository: https://github.com/openclaw/openclaw

Community skill hub: https://clawhub.ai/