TechLatest

Posted on Jun 12 • Originally published at towardsdev.com on Jun 12

OpenClaw Agent Masterclass — Full Tutorial

#opensource #agents #clawdbot #openclaw

Everything you need to install, configure, and extend OpenClaw — the open-source personal AI assistant that runs on your machine and talks to you on the chat apps you already use.

Official home: openclaw.ai · Docs: docs.openclaw.ai · Source: github.com/openclaw/openclaw

This guide follows the product story on the homepage (install → gateway → memory → tools → skills → channels → automation), uses prose and lists only (no comparison tables), and ships terminal + diagram GIFs like our Hermes masterclass.

What you’ll have at the end

OpenClaw installed with the Gateway daemon running
Browser Control UI at http://127.0.0.1:18789/
At least one messaging channel (Telegram recommended for first test)
A configured workspace with SOUL.md and optional ClawHub skill
Understanding of cron , heartbeats , and multi-agent routing

Deploy on OpenClaw VM

Want to skip infrastructure setup?

We provide pre-configured OpenClaw VM images on AWS, Azure, and Google Cloud Platform (GCP). Each deployment comes with OpenClaw, Ollama, and all required dependencies pre-installed, allowing you to launch a production-ready AI agent environment in minutes.

Available with both CPU and GPU configurations for development, testing, and production workloads.

Introduction — the AI that actually does things

OpenClaw is built for a simple promise: message an assistant from your phone, and it does real work on your computer — email triage, calendar checks, shell commands, browser tasks, file edits, and custom workflows via skills.

Unlike a chat-only bot, OpenClaw is self-hosted. Your context, skills, and session history live on your hardware. You pick the model (Anthropic, OpenAI, Google, local Ollama, and more). You control which channels can reach the agent and who is on the allowlist.

Community feedback on openclaw.ai consistently highlights the same strengths: persistent memory, persona onboarding, proactive cron/heartbeats, and the ability to extend the system by chatting (skills, plugins, even prompt hot-reload).

Part 1 — How OpenClaw is structured

OpenClaw centers on one long-running process: the Gateway. It is the control plane for:

Chat channels — WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Matrix, Teams, WebChat, and plugin channels
Agent runtime — tool use, sessions, memory, skills
Control UI — browser dashboard for chat, config, and diagnostics
Companion apps — macOS menu bar, Windows tray, iOS/Android nodes (camera, voice, Canvas)

Docs: Architecture · Gateway

The Gateway is the single source of truth for sessions and routing. CLI commands (openclaw agent, openclaw onboard) and the dashboard all talk to the same core.

Part 2 — Prerequisites

You need Node.js 24 (recommended) or Node 22.19+ for compatibility. OpenClaw fails on older Node versions — if you are stuck on Node 20, use the Node 22 helper from our OpenClaw + Gemma guide.

You also need:

macOS, Linux, Windows 10+, or WSL2
An API key from your chosen provider or a local Ollama install
~5 minutes for onboarding; more if you add WhatsApp or iMessage pairing

Check:

node -v # v22.19+ or v24
which npm

Part 3 — Install

Three paths match openclaw.ai:

One-liner (macOS, Linux, WSL)

curl -fsSL https://openclaw.ai/install.sh | bash

The installer can pull Node and dependencies. On macOS, first run may prompt for Administrator access (Homebrew).

npm global

npm install -g openclaw@latest

Hackable / from source

curl -fsSL https://openclaw.ai/install.sh | bash -s -- --install-method git
git clone https://github.com/openclaw/openclaw.git
cd openclaw && corepack enable && pnpm install
pnpm openclaw onboard

Switch release channels later:

openclaw update --channel stable # or dev

Companion apps (beta): native macOS (15+) and Windows tray apps from openclaw.ai — gateway control, chat, and node features without living in the terminal.

Part 4 — Onboard the Gateway

Run the guided wizard:

openclaw onboard --install-daemon

The wizard walks through:

Gateway bind and authentication
LLM provider and model (API key or Ollama)
Workspace path (default under ~/.openclaw/)
Channel setup (Telegram is the fastest smoke test)
Daemon install (launchd on macOS, systemd on Linux) so the Gateway survives reboots

Onboard wizard — animated

Verify:

openclaw doctor
openclaw gateway status

Part 5 — Open the Control UI

openclaw dashboard

Default URL: http://127.0.0.1:18789/

From the dashboard you can chat, inspect sessions, edit config, and diagnose channel connections. Remote access patterns (Tailscale, SSH tunnel) are documented under Remote access.

Dashboard — animated

CLI chat without the browser:

openclaw agent --message "What can you do on this machine?" --thinking low

Agent CLI message — animated

Part 6 — What lives on disk

After onboarding, OpenClaw owns a home directory. Knowing the layout makes debugging easier.

~/.openclaw/
├── openclaw.json # Main config (channels, models, security)
├── workspace/ # Agent workspace
│ ├── AGENTS.md
│ ├── SOUL.md # Persona / identity
│ ├── TOOLS.md
│ └── skills/ # Installed + custom skills
│ └── <name>/
│ └── SKILL.md
├── credentials/ # Channel tokens (permissions-sensitive)
├── sessions/ # Session metadata
└── … # Logs, cron output, plugin state

openclaw.json is the source of truth for non-secret settings. Secrets and tokens route to appropriate credential stores.

SOUL.md defines who the agent is — tone, boundaries, and behavior. It is the identity layer (similar in spirit to HermesSOUL.md, but living in the workspace).

skills/ is where procedural knowledge lives — bundled skills, ClawHub installs, and agent-authored skills.

Workspace layout — animated

Copy a starter soul from this guide:

cp guides/openclaw/examples/SOUL.md ~/.openclaw/workspace/SOUL.md

Part 7 — Capabilities (from the homepage)

OpenClaw advertises six pillars on openclaw.ai. Here is what each means in practice.

Runs on your machine. macOS, Windows, or Linux. Connect Anthropic, OpenAI, Google, or local models. Data stays on your infrastructure unless a tool explicitly calls an external API.

Any chat app. One Gateway serves many channels. DMs and group chats are supported; group behavior often uses mention rules so the bot does not reply to every message.

Persistent memory. The agent remembers preferences and context across sessions — your assistant becomes specific to you, not a generic chatbot.

Browser control. Navigate pages, fill forms, extract data. Useful for research, booking flows, and admin panels that have no API.

Full system access (configurable). Read/write files, run shell commands, execute scripts. You choose sandbox vs full access based on trust and host environment.

Skills and plugins. Install community skills from ClawHub, add channel plugins, or describe a new workflow in chat and let the agent draft a skill.

Part 8 — Connect messaging channels

Telegram is the quickest first channel: create a bot with @BotFather, paste the token during onboarding or in config.

WhatsApp and iMessage require additional pairing steps documented in the Channels hub.

Minimal allowlist snippet — merge into ~/.openclaw/openclaw.json (full example in examples/openclaw-channels.snippet.json):

{
  channels: {
    whatsapp: {
      allowFrom: ["+15555550123"],
      groups: { "*": { requireMention: true } },
    },
  },
  messages: { groupChat: { mentionPatterns: ["@openclaw"] } },
}

Restart after config changes:

openclaw gateway restart

Security: start restrictive — allowlist phone numbers and require mentions in groups. See Security.

Supported surfaces include WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Google Chat, Matrix, Microsoft Teams, Zalo, WebChat, and plugin channels — 50+ integrations on the marketing site.

Part 9 — Skills and ClawHub

Skills are Markdown with YAML frontmatter — the agent loads descriptions cheaply and pulls full instructions when a task matches.

Install from ClawHub:

openclaw skills search calendar
openclaw skills install <skill-slug>

Browse clawhub.ai. Recent OpenClaw releases emphasize Skill Cards and security scanning (SkillSpector) for hub skills — see the Skill Workshop blog post.

The agent can also author skills from conversation — e.g. “build a skill that checks my WHOOP metrics” — matching patterns described in community shoutouts on the homepage.

Skill Workshop (2026): review and approve proposed skills before they change agent behavior — product direction toward safer self-modification.

Progressive loading keeps token use sane:

Catalog view — names and descriptions only
Full skill — load SKILL.md when triggered
References — optional deep files inside the skill folder

Team-private skills: host a Git repo and install via a slug, using the same pattern as Hermes Skills Hub taps.

Part 10 — Models and local inference

Set or switch models:

openclaw models list
openclaw models set anthropic/claude-sonnet-4
# or local:
openclaw models set ollama/gemma4:e2b

For a full local stack (Ollama + RAG skill), follow OpenClaw + Gemma + RAG.

Providers are swappable without rebuilding the Gateway — the agent runtime handles translation to supported API formats.

Part 11 — Proactive automation: cron and heartbeats

OpenClaw is designed to be proactive , not only reactive.

Cron jobs schedule isolated agent runs — daily briefings, inbox sweeps, reminders. Describe schedules in natural language or use cron syntax. Jobs persist in config and survive Gateway restarts.

Example prompt inside a chat session:

Every weekday at 8am, summarize my calendar and unread priority emails.
Deliver the summary here. Set this up as a recurring cron job.

List jobs:

openclaw cron list

Heartbeats are periodic check-ins — the agent may reach out when something needs attention (community reports surprise check-ins during heartbeats). Configure through workspace and gateway settings per docs.

Useful variants:

One-shot delay: /cron add 30m "Remind me to check the build"
Interval: /cron add "every 2h" "Check server status"
Attach a skill: run a job with --skill so the agent loads a playbook first

Part 12 — Multi-agent routing

One Gateway can route multiple isolated agents — different workspaces, sessions, or senders. Useful for “work agent” vs “personal agent”, or separate Telegram bots.

Concepts:

Session isolation — conversations do not leak context across routes
Workspace per agent — distinct SOUL.mdSkills and tools
Sender-based routing — map channels or users to different agents

Docs: Multi-agent routing

Part 13 — Nodes, voice, and Canvas

Mobile nodes pair iOS/Android apps with the Gateway for camera capture, voice workflows, and Canvas (visual workspace). The macOS/Windows companion apps expose tray controls and local node mode.

Docs: Nodes

This is how users run “fix production from a dog walk” workflows — phone chat triggers agent execution on a home server or Mac mini.

Part 14 — OpenClaw vs Hermes (prose only)

Both are self-hosted, messaging-friendly agent runtimes. Neither is a hosted SaaS.

OpenClaw leads with the Gateway and channels — the product feels like “message your computer from WhatsApp.” Skills extend behavior; the community hub (ClawHub) is large; onboarding and Control UI are polished for personal assistants.

Hermes leads with the learning agent — runtime skill authoring, Curator maintenance, optional GEPA offline validation, and research-oriented tooling (MCP, profiles, training pipeline). See Hermes Agent Masterclass.

You can migrate between them: hermes claw migrate Import OpenClaw-style config into Hermes. Full side-by-side: Hermes vs OpenClaw.

Pick OpenClaw when channel UX, ClawHub, and dashboard-first setup matter most. Pick Hermes when the self-improving skill library and experiment loop matter most. Many operators run one primary runtime and borrow skills from the other ecosystem.

Part 15 — Troubleshooting

openclaw: command not found — reinstall globally or ensure npm global bin is on PATH.

Gateway will not start — runopenclaw doctor; check port 18789 conflicts.

Node version errors — upgrade to Node 22.19+ or 24.

Channel connected but no replies — verify allowlists, mention rules in groups, and bot token.

Model errors — confirm API key in config; test with openclaw agent --message hi.

Docs entry: Troubleshooting

Part 16 — Verify this guide

chmod +x guides/openclaw/scripts/verify-openclaw.sh
./guides/openclaw/scripts/verify-openclaw.sh

Official links

openclaw.ai — product home
docs.openclaw.ai — documentation
github.com/openclaw/openclaw — source
clawhub.ai — skill registry
Discord community

Summary

OpenClaw is a Gateway-first personal agent : install withopenclaw onboard, chat from the dashboard or your favorite messaging app, extend with skills and cron , and keep data on your machine. Start with Telegram and the Control UI, tighten security with allowlists, then add ClawHub skills and automation once the loop feels natural.