✨ Chapter 1
Introduction & Context
Welcome! By the end of this guide you'll have a fully running personal AI agent that can browse
the web, manage your calendar, write code, and respond to you on WhatsApp — all from your own
machine.
What Exactly Is OpenClaw?
OpenClaw is a personal AI assistant that runs on your own devices — your laptop,
phone, or home server — and connects to the messaging apps you already use every day.
🎮
Universal Remote
One interface for your entire digital life
👔
AI Intern
Never sleeps, always ready to act
🔒
Privacy-First
Your data never leaves your devices
🔌
Extensible
Add any capability with Skills
Supported Channels
OpenClaw connects to virtually every major messaging platform: WhatsApp, Telegram, Signal,
iMessage, Slack, Discord, Microsoft Teams, Google Chat, Matrix, Zalo, and a built-in
WebChat UI.
The "Aha!" Moment
You send a WhatsApp message: "Summarize my unread emails and schedule a meeting with Sarah for
Thursday afternoon." OpenClaw:
- Understands your intent via natural language
- Accesses your Gmail via secure API
- Reads your calendar and checks availability
- Generates a summary using an LLM
- Creates the calendar event automatically
- Replies back to you with confirmation ✅
All of this happens on your machine, with your API keys, following your rules.
Why OpenClaw Went Viral
🏠 The Local-First Movement
Your data stays YOUR data — no vendor lock-in, no training on your private information, works offline
for local operations.
🔓 Open Source = Trust + Innovation
Fully open source (MIT License). Audit the code yourself, contribute features, fork and customize
anything — total transparency.
🌉 Chat-First Interface
Instead of juggling 10 apps, just send a message. One interface to rule them all.
🧩 Skills Flywheel
More users → more skills → more value → more users. The ecosystem grows itself.
Evolution Timeline
| Period |
Milestone |
| Nov 2025 |
Peter Steinberger (Austrian vibe coder) publishes Clawdbot —
derived from Clawd (now Molty), named after Anthropic's Claude |
| Jan 27, 2026 |
Renamed Moltbot after trademark complaints from Anthropic; Matt
Schlicht simultaneously launches Moltbook — a social network for AI agents
— sparking viral interest |
| Jan 30, 2026 |
Renamed OpenClaw — "Moltbot never quite rolled off the tongue";
reaches 140,000 GitHub stars & 20,000 forks by Feb 2 |
| Feb 4, 2026 |
ClawCon conference in San Francisco; adopted by companies in
Silicon Valley & China; DeepSeek model integration |
| Feb 14, 2026 |
Steinberger announces joining OpenAI; project moves to open-source
foundation |
Core Theory: Why It Works
🎯 Autonomous Agents
Traditional software is reactive — it waits for input and returns a result. AI
agents are proactive — they plan multi-step tasks, handle errors, and follow up
without being hand-held.
🔄 LLM Routing
| Task Type |
Routed Model |
Reason |
| Coding |
Claude Opus 4.6 |
Best code quality |
| Quick questions |
Gemini Flash |
Speed & low cost |
| Creative writing |
GPT-5.2 |
Creativity |
| Sensitive / offline |
Local Llama |
Privacy |
🧠 The ReAct Pattern
OpenClaw uses the industry-standard ReAct (Reasoning + Acting) loop until the task
is complete:
👁️ Observe
→
🧠 Think
→
⚡ Act
→ repeat
💾 Persistent Memory
| Memory Type |
Description |
Example |
| Short-term |
Current conversation |
"What was I asking about?" |
| Long-term |
Curated knowledge (MEMORY.md) |
"User prefers dark mode" |
| Session History |
Past conversations |
"Yesterday we discussed Python" |
| External |
Files, databases, APIs |
Calendar events, emails |
Cloud AI vs. OpenClaw
| Aspect |
Cloud AI |
OpenClaw |
| Privacy |
Data on vendor servers |
Stays on your devices |
| Offline Use |
❌ Always needs internet |
✅ Works offline |
| Cost |
Subscription fees |
API usage only |
| Model Choice |
Vendor's choice |
Any model you want |
| Control |
Terms of service limits |
You make the rules |
🏛️ Chapter 2
Architecture & High-Level Design
Component 1: The Gateway
The Gateway is the heart of OpenClaw — think of it as an air traffic control tower,
running as a WebSocket control plane on port 18789 by default.
🔗
Connection Hub
WebSocket connections to all channels
📋
Session Manager
Isolated sessions per conversation
⚙️
Config
Reads & hot-reloads openclaw.json
🌐
Web UI
Control panel at /__openclaw__/
bash — key gateway commands
openclaw gateway --port 18789 # Start
openclaw gateway status # Check status
openclaw logs --follow # Stream logs
openclaw config apply # Hot-reload config
Component 2: The Pi Agent Runtime
Pi is OpenClaw's embedded AI processing engine. It handles prompt building, LLM
invocation, tool orchestration, and multi-turn reasoning — all in a real-time streaming loop.
- Streaming responses (real-time output to your channel)
- Block streaming for structured content (code, tables)
- Chaining multiple tools together
- Complex multi-step reasoning
Component 3: Persistent Memory
| File |
Purpose |
When It Loads |
SOUL.md |
Agent personality & voice |
Every session |
USER.md |
User preferences & context |
Main sessions |
MEMORY.md |
Technical notes, lessons, todos |
Main sessions |
AGENTS.md |
Multi-agent rules |
All sessions |
TOOLS.md |
Environment-specific tool configs |
All sessions |
markdown — example SOUL.md
# SOUL.md - Who You Are
You are OpenClaw, a helpful AI assistant. Your traits:
- Professional yet approachable
- Security-conscious and careful
- Always verify before destructive actions
| Category |
Available Tools |
| Core |
Shell exec, file read/write/edit, browser control, Canvas UI |
| Platform |
iOS/Android camera, screen record, GPS location, push notifications |
| Automation |
Cron jobs, webhooks, sub-agents, message send |
| Channel |
Discord reactions/roles, Slack threads/reactions |
directory structure — skills
~/.openclaw/workspace/skills/
├── weather/ # Bundled — SKILL.md
├── github/ # Community — SKILL.md
└── my-custom-skill/ # Yours!
├── SKILL.md
└── helpers.js
🚀 Chapter 3
Getting Started
This section is designed to get you running in under 15 minutes. Follow each step in order and
you'll have a working agent by the end.
Installation
Prerequisites
| Requirement |
Version |
Purpose |
| Node.js |
≥ 22.12.0 |
Runtime |
| npm / pnpm |
Latest |
Package manager |
| Git |
Any |
Skills & plugins |
Install Node.js & OpenClaw
bash — macOS
brew install node
npm install -g openclaw@latest
openclaw --version # confirm install
Run the Onboarding Wizard
bash
openclaw onboard --install-daemon
# Wizard will:
# ✅ Verify prerequisites
# ✅ Create ~/.openclaw/
# ✅ Set up initial config
# ✅ Install system daemon
Verify Everything Works
bash
openclaw doctor
# ✅ Node.js version: v22.x.x
# ✅ Gateway binary: found
# ✅ Config: ~/.openclaw/openclaw.json
Configuration
json — ~/.openclaw/openclaw.json
{
"gateway": {
"port": 18789,
"bind": "loopback",
"auth": { "token": "your-secure-token" }
},
"models": {
"default": "anthropic/claude-4-opus"
},
"channels": {
"telegram": {
"enabled": true,
"botToken": "${TELEGRAM_BOT_TOKEN}",
"dmPolicy": "pairing"
}
}
}
Connect Telegram (Recommended First Channel)
- Message @BotFather on Telegram → send
/newbot → copy the token
openclaw config set channels.telegram.botToken "YOUR_TOKEN"
openclaw gateway --port 18789 --verbose
- Approve your own user:
openclaw pairing approve telegram <code>
Security tip: dmPolicy: "pairing" ensures unknown users get a
pairing code, not AI responses. Never switch to "open" unless in a trusted private
environment.
Model Providers
| Provider |
Setup |
Best For |
| Anthropic ★ |
openclaw login anthropic |
Complex reasoning, code |
| OpenAI |
Set OPENAI_API_KEY |
Creative writing, vision |
| Google Gemini |
Set GEMINI_API_KEY |
Fast & cheap queries |
| Ollama (local) |
localhost:11434 |
Sensitive / offline |
json — model failover
{
"models": {
"default": "anthropic/claude-4-opus",
"fallbacks": ["openai/gpt-5", "ollama/llama3.3"]
}
}
Skills
Skills are Markdown files that teach OpenClaw how to use tools. They load automatically at startup.
| Skill |
What it does |
weather |
Fetch weather forecasts |
github |
Interact with GitHub repos |
browser |
Control a web browser |
canvas |
Render visual UIs |
cron |
Schedule recurring tasks |
bash — install & create skills
clawhub install todoist # install from marketplace
# Create your own skill
mkdir -p ~/.openclaw/workspace/skills/greeting
cat > ~/.openclaw/workspace/skills/greeting/SKILL.md << 'EOF'
---
name: greeting
description: Generate personalized greetings
---
# Greeting Skill
Greet users warmly by name, matching time of day.
EOF
💡 Chapter 4
Practical Applications & Productivity
Best Use Cases
🏆 Absolute Game-Changers
| Use Case |
Example Command |
| Daily Standups |
"Generate my standup for today" |
| Email Triage |
"Summarize unread emails, flag urgent" |
| Research |
"Research React 19 features and summarize" |
| Code Review |
"Review this PR: github.com/…" |
| Documentation |
"Document the functions in this file" |
🥈 Excellent Productivity Boosts
| Use Case |
Example Command |
| Meeting Prep |
"Prep me for my 2pm with Sarah" |
| Travel Planning |
"Plan a 3-day trip to Tokyo" |
| Content Creation |
"Write a blog outline about AI agents" |
| Bug Analysis |
"Analyze this error log and suggest fixes" |
Daily Workflows
Morning Routine
conversation example
You: "Good morning!"
OpenClaw:
→ Weather: 72°F, sunny
→ Calendar: 3 meetings (10am, 2pm, 4pm)
→ Emails: 12 new, 2 flagged as urgent
→ Reminder: Submit expense report by EOD
You: "Block 2 hours for deep work between meetings"
OpenClaw: ✅ Created: Focus Time 12:00–2:00 PM
You: "Summarize urgent emails and draft responses"
OpenClaw: [Short summaries + reply drafts]
Research & Analysis
conversation example
You: "Research LLM efficiency in 2025. Find 3 papers, summarize approaches."
OpenClaw:
1. Opens browser → searches arXiv
2. Reads abstracts & key sections
3. Synthesizes findings
→ Detailed comparison with citations & links
Building Your Personal Assistant Personality
markdown — SOUL.md
# SOUL.md
You are my personal productivity assistant.
## Style
- Concise but warm
- Lead with key information
- Use bullet points for clarity
## Proactive Behaviors
- Remind me of deadlines 24h in advance
- Flag potential meeting conflicts
Scheduled Automation (Cron Jobs)
bash
# Morning briefing at 8am every day
openclaw cron add --name "morning-briefing" \
--schedule "0 8 * * *" \
--message "Generate morning briefing with weather and calendar"
# Weekly review every Friday 4pm
openclaw cron add --name "weekly-review" \
--schedule "0 16 * * 5" \
--message "Weekly review: What did we accomplish this week?"
🛡️ Chapter 5
Security & Reliability
Model Selection for Performance
| Task |
Recommended Model |
Why |
| Quick questions |
Gemini Flash / GPT-4o-mini |
Fast & cheap |
| Complex reasoning |
Claude Opus / GPT-5 |
Deep thinking |
| Code generation |
Claude Opus |
Best code quality |
| Sensitive data |
Local Llama / Qwen |
Privacy |
The Threat Model
| Risk |
Severity |
Mitigation |
| Prompt Injection |
🔴 Critical |
Use resistant models (Claude), scan patterns |
| Malicious Skills |
🔴 Critical |
Review code, use allowlists |
| Dangerous Tools |
🔴 Critical |
BlockList, require confirmation |
| Untrusted DMs |
🔴 Critical |
Always use dmPolicy: "pairing" |
| Gateway Exposure |
🔴 Critical |
Bind to loopback, never 0.0.0.0 |
json — critical security settings
{
"channels": {
"telegram": { "dmPolicy": "pairing", "allowFrom": [] }
},
"tools": {
"exec": {
"blockList": ["rm -rf", "sudo", "dd"],
"requireConfirmation": true
}
},
"gateway": { "bind": "loopback" }
}
Known Challenges & Mitigations
| Challenge |
Current Mitigation |
| Hallucination |
Tool verification, confidence scoring |
| Context Limits |
Session pruning, RAG |
| Cost Management |
Model routing, usage tracking |
| Reliability |
Retry policies, automatic fallbacks |
🎓 Conclusion
You're Ready to Build
"The best way to predict the future is to build it." — Go build something amazing with OpenClaw!
🚀
Key Takeaways
✅
Local-First — runs on your devices, data stays private
✅
Architecture — Gateway → Pi Runtime → Tools → Skills
✅
Memory — SOUL.md, USER.md, MEMORY.md define your agent
✅
ReAct loop — Observe → Think → Act, until task complete
✅
Skills extend — ClawHub marketplace for community plugins
✅
Security — pairing mode, loopback binding, tool blocklists
Your Next Steps
- Install:
npm install -g openclaw@latest
- Configure one channel (Telegram is the easiest to start)
- Set up your SOUL.md and USER.md personality files
- Try a few basic commands and get comfortable
- Install 2–3 skills from ClawHub
- Build one automation — a morning briefing is a great first project!
Resources