Video Summary - OpenClaw Full Tutorial for Beginners – How to Set Up and Use OpenClaw (ClawdBot / MoltBot)

OpenClaw Full Tutorial — Summary

Overview

OpenClaw (formerly Claudebot / MoltBot) is a self-hosted, proactive autonomous agent runtime that connects messaging platforms (WhatsApp, Telegram, Discord, etc.) to agents which can execute real-world tasks such as calendar/email triage, running shell commands, controlling smart devices, and posting to social media.

Course goal: install and run OpenClaw, connect models, manage persistent memory, add and author skills, create multi-agent setups, and secure/sandbox agents so you can operate a persistent 24/7 AI operator while keeping data under your control.

Prereqs & Environment

Primary interface: CLI (TUI is also used). Familiarity with terminal commands expected.
Recommended knowledge: LLM basics, API usage, prompt/context engineering.
Node.js required: version >= 22 (examples use Node 25).
Hosting options:
- Self-host locally (full host access) — useful for desktop/browser integrations.
- VPS recommended for isolating host files and reducing risk.

Installation & Onboarding (practical steps & defaults)

Install via npm:
- npm install -g openclaw
Run the onboarding wizard:
- openclaw onboard
- Use --install-daemon (or install-daemon flag) to install the gateway as a background service (launchd/systemd) so it auto-starts.
Defaults:
- Workspace path: ~/.openclaw (config and agent files live here).
- Default gateway port: 18789.
- Gateway bind choices: loopback (recommended for single-machine), LAN, Tailscale, auto/custom IP — changeable later via CLI.
Model choices: Anthropic, Google, etc. Anthropic is recommended in the demo. Note: Google Flash 3 has a small free quota (20 req/day).
Gateway token and channel pairing are handled during onboarding (CLA setup token / secret keys).

Core Concepts

Gateway: long-running process that maintains platform connections and routes incoming messages to agents.
Workspace: directory containing agent definitions, identity.md, agent.md, heartbeat.md, soul/tools scratch-pad, and user metadata. These files are editable and can be backed up to Git.
Memory / Persistence:
- bootstrap.md for one-time initialization.
- Heartbeat runs periodically (default hourly) to perform recurring tasks.
Configuration: JSON config files exist, but using CLI/TUI is safer than manual edits.

TUI / CLI Features

Terminal User Interface (TUI): interactive setup and agent chat; supports many slash commands (around 46 defaults).
Health & security commands:
- openclaw security-audit
- openclaw doctor (health checks/fixes)
- openclaw status / openclaw health
Slash commands: switch agents, run skills, inspect gateways, and more.

Channels & Demo Examples

Pinchboard (example social app built with OpenClaw): demonstrates agent registration, verification via tweet, and posting (“pinch”).
WhatsApp:
- Enable plugin and pair: openclaw plugins enable whatsapp, then openclaw channels login to scan QR and pair account.
- Recommended: use a dedicated phone number for the bot.
Discord:
- Create application, reset bot token, enable message content intent, invite bot with required scope.
- Add guild/channel IDs to OpenClaw config so the bot can respond.
Email:
- Example “send email” skill written in Python using SMTP.
Multi-channel demos showed web chat, WhatsApp, Discord, and email operating together.

Skills & Extensibility

Skills:
- Markdown-based (skill.md) with YAML front matter that instructs agents how to use tools/APIs and teach workflows.
- Examples: Obsidian skill, Apple Notes via memo CLI, Gmail/email via Himalaya, image tools, Twitter, etc.
Organization:
- Per-agent skills live in the agent’s skills folder.
- Shared skills live in openclaw/skills and are visible to all agents.
Invocation:
- Skills can be user-invocable (slash commands).
- Token impact: each skill adds characters to the system prompt (roughly ~24 tokens per skill, varies).
Clawhub:
- Skill registry at clawhub.com.
- Install via clawhub install <skill> (conceptually analogous to pip).
- Treat third-party skills as untrusted — review before enabling.

Multi-agent Support

Create and manage multiple agents with separate workspaces, profiles, tokens, sandboxes, and tool policies (for example: separate “work” and “personal” agents).
CLI commands:
- openclaw agents add <name>
- openclaw agents list
In the TUI, use /agents to switch active persona/session — useful to separate permissions, integrations, and behaviors between contexts.

Security & Sandboxing (critical)

Main risks:

Prompt injection and arbitrary command execution — an agent with host access can run shell commands and modify files.

Best practices:

Run on a VPS or use Docker sandboxing to isolate the host file system.
Keep channels restricted (sender allow lists); avoid placing bots in group chats.
Use a dedicated phone number for bot accounts.
Scope and rotate tokens; treat third-party skills and code as untrusted.
Run openclaw security-audit and openclaw doctor regularly.
Prefer stronger LLMs (less susceptible to prompt manipulation).

Sandboxing modes and scope:

Modes: none, all, sandbox (controls whether containers are used).
Scope:
- session — new container per session,
- agent — one container per agent,
- shared — all sandboxed agents share a single container.
Tool policies:
- Use tool.deny to block risky tools (e.g., exec, process, browser).
- Elevated exec runs bypass the sandbox and execute on the host — dangerous and should not be enabled for untrusted senders.
Browser control: restrict allowed domains or use sandboxed browsers.
Demo: a Docker-based sandbox for the “work” agent (agent scope) was created and verified to be unable to access host desktop files.

Practical Tips & Commands Referenced

Install globally:
- npm install -g openclaw
Onboard:
- openclaw onboard (use install-daemon flag to install gateway service)
Channels login (WhatsApp):
- openclaw channels login
Enable plugin:
- openclaw plugins enable whatsapp
Agent management:
- openclaw agents add <name>
- openclaw agents list
Security / health:
- openclaw security-audit
- openclaw doctor
- openclaw status
- openclaw health
Skills registry:
- clawhub install <skill>
Backup:
- Keep workspaces backed up via Git (repo docs include instructions).
Defaults and paths:
- Workspace and token locations are created under ~/.openclaw (or a path chosen during onboarding).

Tutorial Structure (modules covered)

Recap of what OpenClaw is
Installation and onboarding
Workspace and memory (agent definition files)
Pinchboard (social agent demo)
Building a personal assistant with channels (WhatsApp, Discord, email)
Skills (writing, installing, invoking; Clawhub)
Multi-agent setups (separate workspaces/permissions)
Security (threats and mitigations)
Sandboxing (Docker isolation and scope options)

Example Demos Shown

Onboarding and naming an agent (agent named “Nova”).
Pinchboard: agent account creation and tweet verification.
WhatsApp integration: QR pairing and 1:1 message exchange.
Discord integration: bot creation, tokens, adding to server, responding to mentions, and demonstrating host file access.
Email skill: send_email.py skill using SMTP to send an email.
Created a sandboxed “work” agent (Docker) and verified it could not access host files.

Recommendations Emphasized

Start conservative: limit channel access and tool permissions.
Use sandboxing or run on a remote VPS to protect important files.
Review third-party skills before enabling them.
Use the CLI/TUI for configuration changes instead of editing JSON files directly.
Back up agent workspaces to Git to persist configurations across devices.

Primary Speaker / Sources

Instructor: Kian — presenter of the course and demos.
Project: OpenClaw software/repository (referenced GitHub repo and GitHub Pages docs).
Models & registries referenced: Anthropic, Google Flash, the Claude family, and the Clawhub skill registry.

End of summary.

OpenClaw Full Tutorial for Beginners – How to Set Up and Use OpenClaw (ClawdBot / MoltBot)

Key takeaways