How to Use OpenClaw: Setup, Skills, and Automations That Work

In November 2025, Austrian developer Peter Steinberger pushed a project called Clawdbot to GitHub. By March 2026, it had accumulated more GitHub stars than any software project in history — over 250,000 — survived two forced renamings, attracted enterprise bets from NVIDIA and Tencent, and prompted Jensen Huang to ask a room full of developers at GTC: “What’s your OpenClaw strategy?” That’s not hype. That’s a real signal that something structurally different is happening here. OpenClaw is an open-source, locally-running AI agent framework that uses your existing messaging apps — WhatsApp, Signal, Telegram, iMessage, Discord — as the interface. You bring your own API key, point it at whatever model you want, and extend it through a skill system that’s already grown to 100+ built-in capabilities. This guide is about how to actually use it: setup, skills, real automations, and the security considerations you can’t afford to skip.

What OpenClaw Actually Is (and Why the Messaging Interface Matters)

Most AI tools give you a new interface to learn. A chat window, a dashboard, a sidebar in your IDE. OpenClaw does the opposite — it lives inside the apps you’re already in 40 times a day. You message your OpenClaw agent the same way you’d message a friend on WhatsApp or Signal, and it responds, takes actions, and runs automations from there.

This sounds like a small UX detail. It isn’t. The reason people don’t use AI tools consistently isn’t capability — it’s friction. When your agent lives in iMessage, you don’t have to remember to open a separate app. You just message it when you think of something. That behavioral shift is why OpenClaw spread the way it did: 60,000 GitHub stars in the first 72 hours after it went viral, which tells you something real about pent-up demand for this kind of ambient, always-available agent.

The architecture is worth understanding briefly. OpenClaw runs locally on your machine. It’s not a cloud service. You connect it to a messaging platform, give it access to a language model (Claude, GPT-4o, Gemini, DeepSeek, or a local model via Ollama), and extend its capabilities through skills — which are just directories with a SKILL.md file that tells the agent what the skill does and how to use it. That’s the whole system. Simple enough that a developer can fork it and add skills in an afternoon. Powerful enough that NVIDIA built an enterprise layer called NemoClaw on top of it.

The mascot is a red lobster. The tagline is “The lobster way.” Steinberger has a sense of humor about the whole thing — this started as a hobby project, got named Clawdbot, then got a trademark complaint from Anthropic (presumably over the “claw” association with Claude), got renamed to Moltbot on January 27, 2026, then renamed again to OpenClaw three days later. The community largely found this funny. The 250,000 stars kept coming anyway.

Setup: From Zero to Running Agent

OpenClaw is MIT licensed and free to use. The only ongoing cost is your API usage — whatever you pay for Claude, GPT-4o, or whichever model you connect. If you want to run fully locally and free, Ollama integration means you can point it at a local model like Llama 3 or Mistral.

The basic setup path looks like this:

1. Clone the repo from GitHub. The main repository is at the OpenClaw GitHub org. Installation is documented in the README — Node.js based, standard dependency install.
2. Configure your model. Add your API key for whichever provider you’re using. OpenClaw supports Claude (Anthropic), GPT models (OpenAI), Gemini (Google), DeepSeek, and local models via Ollama. You set this in the config file.
3. Connect your messaging interface. This is where most first-time users spend the most time. Each platform has different connection requirements. WhatsApp requires a WhatsApp Web session. Telegram and Discord have bot API flows. iMessage works on Mac via the Messages app bridge. Signal is the most technically involved due to Signal’s architecture.
4. Test with a basic prompt. Message your agent something simple — ask it to summarize a URL, set a reminder, or tell you the weather — to confirm the pipeline is working end-to-end.
5. Install skills from ClawHub. The ClawHub registry is the community skills marketplace. Browse it, find skills relevant to what you actually do, and drop the skill directories into your skills folder.

One honest note on setup complexity: this is not a one-click install. It’s a developer-grade tool. If you’re comfortable with a terminal and have worked with APIs before, you’ll be up in under an hour. If you’re not, expect to spend meaningful time on the messaging platform connection step specifically. The community Discord is active and helpful.

The Skills System: How OpenClaw Extends Itself

Skills are what make OpenClaw genuinely useful rather than just a fancy chatbot wrapper. Each skill is a directory containing at minimum a SKILL.md file — a plain-text description of what the skill does, what inputs it takes, and any setup it requires. The agent reads these files and knows how to invoke the skill when relevant.

There are 100+ built-in skills covering common automations: web search, calendar access, file operations, email drafting, reminders, weather, note-taking, and more. The ClawHub registry extends this significantly with community-built skills.

Skills worth knowing about (based on what the community actually uses most):

• Web research skills — Search, scrape, and summarize web content on request. Useful for quick competitive research, news monitoring, or just answering questions without switching apps.
• Calendar and scheduling skills — Connect to Google Calendar or iCal, check availability, draft meeting summaries, set reminders.
• File and document skills — Summarize PDFs, extract data from documents, generate drafts based on templates.
• Code skills — Run code snippets, explain errors, generate boilerplate. Developers use this constantly in their terminal workflow via the messaging interface.
• Communication skills — Draft emails, summarize long threads, generate replies in your style.
• Data skills — Parse CSVs, run calculations, generate quick reports from structured data.

Building a custom skill is genuinely approachable. If you have a workflow that involves an API — pulling data from Notion, querying your CRM, checking a specific internal dashboard — you can write a SKILL.md that describes what the skill does and a small script that executes it. The agent handles the rest: knowing when to call it, how to parse your natural language request into the right inputs, and what to do with the output.

One important caveat: the ClawHub skills registry had a documented security compromise (more on this below). Be selective about skills you install from third-party sources and review what access each skill requests before deploying.

Real Automations That Actually Work

The difference between a tool you use once and a tool that changes your actual workflow is specificity. Here are concrete automation patterns that work well with OpenClaw in practice. If you want to see how OpenClaw stacks up against other AI agent tools before going deeper,