CodeAdvancedOpenClaw

OpenClaw: The Open-Source AI Agent That Lives on Your Machine

OpenClaw is the viral 2026 open-source personal AI agent — 250K+ GitHub stars in 60 days. It runs locally, talks through Telegram/Signal/Discord/WhatsApp, executes shell commands, and uses Claude or GPT under the hood.

May 15, 2026·5 min read
Share:
OpenClaw: The Open-Source AI Agent That Lives on Your Machine

What is OpenClaw in 2026?

OpenClaw is a free, open-source autonomous AI agent that runs on your own machine and uses messaging apps as its main interface. Unlike chatbots that only talk, OpenClaw has eyes and hands: it can read and write files, run shell commands, browse the web, and call APIs on your behalf — all while you chat with it from your phone via Telegram, Signal, Discord, or WhatsApp.

It went viral in early 2026: by March, the GitHub repo had crossed 250,000 stars, briefly overtaking React as the most-starred project on GitHub. The project was created by Austrian developer Peter Steinberger as "Clawdbot" in late 2025, renamed to "Moltbot" after a trademark complaint from Anthropic, then renamed again to OpenClaw a few days later. Steinberger has since joined OpenAI, and the project continues as an active community-driven open-source codebase.

Why OpenClaw Matters

Most AI products are SaaS: your data goes to a vendor, you pay a monthly fee, you can't extend it. OpenClaw flips this. The agent runs locally on macOS, Linux, or Windows (via WSL2), reads its config from ~/.openclaw/openclaw.json, and connects to whichever LLM you already pay for — Claude, GPT, DeepSeek, or a local Ollama model. Skills are just folders. The codebase is hackable. You can read every line of what the agent does.

The other key idea is persistence. OpenClaw doesn't just respond to messages — it runs in the background on a heartbeat, evaluates its task list, and acts. That makes it the closest thing the open-source world has to the agent OpenAI and Anthropic keep promising in their roadmaps.

Installation in 60 Seconds

The fastest path:

npm install -g openclaw@latest
openclaw onboard --install-daemon

The onboard wizard walks you through the gateway, workspace, channels, and skills. Node 24 is recommended (22.16+ minimum). On macOS, install the companion menu-bar app for voice wake and push-to-talk. On Windows, run inside WSL2 — it's officially supported but not native.

For the source install:

git clone https://github.com/openclaw/openclaw.git
cd openclaw && corepack enable && pnpm install
pnpm openclaw setup
pnpm gateway:watch

Picking a Model

Configure your model in ~/.openclaw/openclaw.json:

{
  "agent": {
    "model": "anthropic/claude-opus-4-7"
  }
}

Recommended pairings in 2026:

  • Claude Opus 4.7 — Best for agentic work. Long-horizon planning, careful shell usage, fewer "I'll just delete your inbox" moments.
  • Claude Sonnet 4.6 / 4.8 — Cheaper, still excellent. Default for high-volume background tasks.
  • GPT-5.5 — Strong all-around, slightly faster turn times via the OpenAI Responses API.
  • DeepSeek V4 — Cheapest serious option. Good for skill development where API costs add up fast.
  • Local (Ollama) — Privacy-first. Slower and weaker, but everything stays on your laptop.

Channels: Talk to Your Agent From Anywhere

OpenClaw exposes itself through 20+ messaging platforms — WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, IRC, Microsoft Teams, Matrix, Feishu, LINE, Mattermost, Twitch, WeChat, and more. The most common setup is Telegram or Signal for personal use, Slack or Discord for shared/team agents.

By default, unknown senders are blocked. Anyone who DMs your bot first receives a pairing code; you approve them explicitly:

openclaw pairing approve telegram a3f2k9

This is intentional — opening DMs to the public via dmPolicy="open" turns your agent into a target for prompt-injection attacks.

Skills: How OpenClaw Actually Does Things

Skills are small folders inside ~/.openclaw/workspace/skills/<skill>/SKILL.md. Each defines a capability — read email, post to GitHub, control Hue lights, query a database. OpenClaw ships with 100+ built-in skills, and ClawHub is the community registry where you can pull more or publish your own.

Popular skill categories:

  • System — bash, file read/write, edit, process control (enabled by default in your main session)
  • Browser — Playwright-based web automation (disabled by default; opt in per session)
  • Productivity — Gmail, Calendar, GitHub, Obsidian, Linear, Notion
  • Smart home — Philips Hue, HomeKit, Home Assistant
  • DevOps — Docker, kubectl, AWS, Vercel, GitHub Actions

The killer feature: the agent can write its own skills. Ask it to "build me a skill that summarizes my Linear tickets every morning" and it will scaffold the SKILL.md, generate the code, and reload itself.

Sandboxing & Security

OpenClaw runs with full system access in your main session — that's the point. But you should never let untrusted input near that session. For group channels, set:

"agents": {
  "defaults": {
    "sandbox": { "mode": "non-main" }
  }
}

Non-main sessions run inside Docker with browser, canvas, cron, and external integrations denied by default. Treat every inbound DM as untrusted. Read the Security and Sandboxing docs before exposing anything to the internet — community skills have already shipped with credential-stealing payloads, and an agent with shell access is a juicy target.

Five Workflows People Actually Run

  • Inbox triage — Every hour, classify new email, draft replies for routine threads, surface only what needs you. Combine the Gmail skill with a Claude Sonnet model for cost control.
  • Calendar concierge — Reschedule meetings, book travel, send confirmations. Pair with the WHOOP/Health skills for "don't book before 9am after a red recovery day."
  • Code review buddy — Watch a GitHub repo, summarize new PRs in your Telegram, auto-generate Claude Code review comments on draft PRs.
  • Personal CI/CD — Run nightly builds, post diffs to Discord, auto-revert on regression. Cheaper than a CI provider for solo developers.
  • Home assistant — Reset Hue lights at sunrise, monitor air purifier filters, ask the agent "what's running" via Signal voice notes.

The Catch

OpenClaw is not for everyone. The setup curve is real: you need a comfortable command line, a paid LLM API, and the discipline to read what skills do before installing them. It also burns through tokens — autonomous agents that run on a heartbeat can quietly rack up $30-$100 a month even at Sonnet prices. Monitor your dashboards.

If those tradeoffs are fine with you, OpenClaw is the most powerful personal AI setup you can build in 2026 without writing the orchestration layer yourself. The repo is at github.com/openclaw/openclaw; start there, run openclaw doctor if anything looks off, and join the Discord for skill help.