The build route through OpenClaw

OpenClaw is not a complex system to set up, but it is not a single command install either. There are choices to make, external credentials to arrange, and a sequence of steps each deserving its own check. What follows is a hands on route, not reference documentation. Follow it in order, do not skip, and use the checks after each step to confirm you are on track.

01 · WHAT YOU NEEDThe prerequisites on your machine

Before you start, the host has to be ready. The list is short but every item is a hard requirement.

• OS: macOS (best tested), Linux, or Windows. For the macOS menu bar app you need macOS; for headless servers, Linux.
• Node.js 22 or higher. OpenClaw is TypeScript and runs on Node. Check with node --version. Install if needed via Homebrew (brew install node@22) or a Node version manager.
• At least one LLM API key. Anthropic Claude is the recommended first choice (ANTHROPIC_API_KEY). Alternatives: OpenAI, Gemini, OpenRouter, or local via Ollama. For local Ollama you do not need an API key.
• Enough disk and RAM. OpenClaw itself is light; the heavy lifting sits in Docker images for sandboxing and any local models. Allow at least 4GB RAM and 5GB disk.
• Docker, only if you want sandboxing turned on. Default sandbox mode is non-main, so plan for Docker unless you disable it.
• Network. Outbound HTTPS to your LLM provider and messaging platforms. No inbound needed unless you offer hooks via webhooks.

02 · INSTALLATIONPutting OpenClaw on your machine

The installation is one command followed by a verification step.

npm install -g openclaw@latest

Verify:

openclaw --version

You should see a version like 2026.5.22. If the command is not found, add the npm prefix path to your shell rc:

npm prefix -g
# add the printed path to ~/.zshrc:
export PATH="$(npm prefix -g)/bin:$PATH"

On macOS a sharp related install error can occur. In that case try:

SHARP_IGNORE_GLOBAL_LIBVIPS=1 npm install -g openclaw@latest

03 · FIRST SETUPThe onboarding wizard

With OpenClaw installed, run the onboarding wizard. It walks you through the essential configuration choices.

openclaw onboard --install-daemon

The wizard runs through four steps.

Gateway authentication. You get the option to set a token to secure the gateway. Generate one with openssl rand -hex 32 and put it in an environment variable. Do not hardcode tokens in openclaw.json; use ${OPENCLAW_GATEWAY_TOKEN} syntax.

LLM provider. Enter your API key for the provider you chose. Anthropic wants ANTHROPIC_API_KEY, OpenAI wants OPENAI_API_KEY, and so on. For local Ollama no key is required; OpenClaw detects Ollama automatically on localhost:11434.

First channel. Here you can add a first channel or skip and add one later. For your first time, skipping is fine; learn to talk to your claw via the TUI or the macOS menu bar app first, then connect a channel.

Daemon install. On macOS via launchd, on Linux via systemd. This makes the gateway run in the background without keeping a terminal open.

After this:

openclaw status

Shows running status, connected providers, and any warnings.

04 · YOUR FIRST AGENTTalking to a claw

With the gateway running you can talk to your claw for the first time. Three ways to start:

Via TUI (quick, in the terminal):

openclaw tui

Via the macOS menu bar app (on macOS): click the OpenClaw icon in your menu bar and open a chat window.

Via the web dashboard:

openclaw dashboard

Or navigate directly to http://127.0.0.1:18789/.

Send a test message: “Hello, who are you?” The claw should reply with its name (default is something like Claw) and a short introduction. If that does not work, check openclaw doctor:

openclaw doctor --fix

That tries to resolve configuration issues automatically. For manual diagnosis:

openclaw logs --follow

Follows the logs live. Errors look explicit, usually with a suggestion for the fix.

05 · ADDING SKILLSGiving the claw new abilities

Skills are the abilities of the claw. Without skills it can talk and use a few built in tools; with skills it becomes useful for specific domains.

See which skills are already available:

openclaw skills list
openclaw skills list --eligible

An “eligible” skill is one whose prerequisites (such as API keys or external tools) are all present. For skills that need specific external software:

openclaw skills check --verbose

Shows per skill what is missing. Install missing dependencies where needed.

Three first skills that are usually immediately useful:

• memory-search: semantic search across the claw’s own memory. Enable via agents.defaults.memorySearch.enabled: true in config.
• web-search and web-fetch: the claw can go on the web. For web-search an API key for a search provider is usually needed.
• apple-notes (on macOS): access to Apple Notes. Useful right away for anyone already organising there.

Writing custom skills happens in a folder inside your workspace:

mkdir -p ~/.openclaw/workspace/skills/my-skill

Create a SKILL.md:

---
name: my_skill
description: "What the skill does"
metadata:
  openclaw:
    emoji: "🔧"
    requires:
      bins: ["curl"]
      env: ["MY_API_KEY"]
---

# My Skill

When the user asks for X, do Y, using Z.

Hot reload picks up new or changed skills within 250ms; no restart needed.

06 · CHANNELS AND HOOKSThe claw in your messaging apps

With a working claw and relevant skills, connect a channel. Start with one; add more only when the first is working.

Telegram is the safest first choice (official Bot API):

1. Create a bot via @BotFather and note the token.
2. Add the token to your environment:

   export TELEGRAM_BOT_TOKEN="123456:ABC..."
   

3. The gateway picks it up via hot reload. Verify:

   openclaw channels status
   

Discord takes a bit more: create a bot in the Discord Developer Portal, turn on Message Content Intent, and add the token.

WhatsApp is possible but risky. Read Part 3 of the fieldguide first: WhatsApp forbids unofficial clients in its ToS and can ban your number. Never use your main number; use a second SIM or virtual number.

Hooks are connected via:

openclaw webhooks gmail setup --account your-email@gmail.com --project gcp-project

For other hooks: configure hooks.mappings in openclaw.json. A simple hook lives in ~/.openclaw/hooks/<name>/ with HOOK.md and handler.ts.

07 · DEPLOYMENTFrom your laptop to something more durable

For personal use your laptop is enough. But the claw becomes really valuable only when it keeps running while your laptop is closed. Three deployment paths, in increasing complexity.

Always on, on your own machine. The simplest option. With openclaw onboard --install-daemon the gateway already runs as a background service. Make sure your laptop does not sleep, or accept that the claw only lives when you are awake.

On your own server. A Mac Mini, a Linux VPS, or a local Raspberry Pi 5. Install OpenClaw, configure with your credentials, run as a service. Bind the gateway to loopback if you reach it through SSH tunnels or Tailscale; to lan if the server is on your network and you want to connect from other devices.

Containerised. A Docker image for OpenClaw itself, optionally with orchestration. The fieldguide does not go deep here; the pattern is the same as any Node process in production. Pay attention to: persistent volumes for the workspace, environment variables for credentials, network policy for inbound webhooks.

For production grade setups: monitor the gateway via openclaw status --deep, integrate with your log aggregation, configure cron failure alerts.

08 · MAINTENANCEThe recurring tasks

OpenClaw is not install and forget. A few recurring tasks:

Updates. Weekly or on major fieldguide updates:

openclaw update

Pulls the newest version, validates config compatibility, restarts the gateway. For bleeding edge: --channel beta. For always stable: nothing, the default is stable.

Security audits. After every update:

openclaw security audit
openclaw security audit --fix

Runs the six security checks (gateway bind, auth, DM policy, file permissions, tool restrictions, mention policy). --fix resolves what it can automatically.

Memory maintenance. Weekly or when memory gets messy:

openclaw memory status
openclaw memory index

For anyone using the optional memory wiki:

openclaw wiki doctor

Cron checks. Inspect running cron jobs and their status:

openclaw cron status
openclaw cron list

Backups. Once your workspace has become substantively valuable:

openclaw backup create

Makes a tar of workspace, sessions and config. Keep it in a safe place; this tar holds no credentials but does hold your personality files and conversation history.

09 · NEXT STEPSWhere you can go from here

The build route ends here, but OpenClaw begins here. Three directions for what to do next.

Multiple agents and multi channel setups. Work and home as separate claws, each with their own workspace, personality, model. Routing rules in openclaw.json decide who serves which channel. Start with two agents and expand once you know what they should know about each other.

Heartbeat and cron for proactivity. Make the claw responsible for recurring tasks. A daily briefing at 7:00, a weekly project review on Monday, an open issues check every hour. Write the rules in HEARTBEAT.md or via openclaw cron add.

Writing your own skills. When you see usage patterns not covered by existing skills, write your own. A skill is a folder with a markdown file and possibly scripts. Hot reload makes iterative writing pleasant.

For the conceptual side of OpenClaw, read the Mental Pill on this body. For the technical depth, the Developer Pill. For the risk layer, the Security Pill. For the source itself without a lens, the Placebo.