TL;DR
- OpenClaw is a free, open-source AI agent that plugs into WhatsApp, Telegram, or Discord — so instead of opening a separate app, you just text your AI like a contact.
- It was built in a few weeks by Austrian developer Peter Steinberger, hit 247,000 GitHub stars in under four months, and led to Steinberger being hired by OpenAI to build the next generation of personal agents.
- China has already surpassed the US in adoption, launched government subsidy programs around it, and coined a phrase for setting it up: “lobster farming.”
Sometime in late 2025, an Austrian developer named Peter Steinberger shipped a small open-source project to GitHub. He called it Clawdbot — a nod to Anthropic’s Claude chatbot, which was powering it under the hood. Within four months, the project had a quarter million GitHub stars, a trademark dispute with Anthropic, a new name twice over, and a job offer from Sam Altman. China built an entire cottage industry around it. The concept it demonstrated — a personal AI agent that lives inside the chat app already on your phone — is now one of the most-copied ideas in tech.
This is the story of OpenClaw: what it is, where it came from, and how to get it running in about ten minutes.
What OpenClaw actually does
The core idea is simple and, in retrospect, obvious. Every major AI lab has a web interface. Most also have APIs. But people spend their time in messaging apps — WhatsApp, Telegram, Discord, iMessage. OpenClaw is a bridge between those two worlds. You install it on a server or your local machine, connect it to an LLM provider of your choice, link it to your Telegram account or WhatsApp number, and from that point on your AI agent is just a contact in your phone.
You can ask it to search the web, summarize documents, write code, draft emails, manage files, or run multi-step tasks while you’re away from your computer. Because it stores configuration and memory locally, it actually remembers context across conversations — a limitation that trips up most web-based AI tools. It supports Claude, GPT-4o, DeepSeek, Gemini, and most other major LLM providers through a unified interface. You pick the brain; OpenClaw handles the plumbing.
As of early 2026, OpenClaw supports 12 messaging channels simultaneously — Telegram, WhatsApp, Discord, Slack, iMessage, Signal, and more. You can run all of them at once, fielding the same AI agent across every chat app on your phone.
From Clawdbot to Moltbot to OpenClaw: the naming saga
Steinberger published the first version on November 24, 2025 under the name Clawdbot. The name was intentionally close to Claude — that was the whole point, it was a bot built on Claude. Anthropic’s legal team apparently did not share his sense of playfulness. On January 27, 2026, following trademark complaints, Steinberger renamed the project Moltbot, staying on theme with a lobster-molting metaphor. Three days later, he renamed it again because Moltbot “never quite rolled off the tongue.” OpenClaw stuck.
The trademark dispute was brief, and somewhat ironic given that Anthropic had nothing to do with the project’s success. Anthropic’s own Claude Code — its professional coding agent — was a separate product aimed at developers. What Steinberger built was something Anthropic hadn’t: a zero-friction agent for ordinary people, running in the app they were already using, costing nothing to try.
By the time the name settled on OpenClaw, the GitHub star count was already in the six figures. Anthropic eventually responded with Claude Code Channels, a feature that lets users connect Claude Code to Telegram and Discord — essentially the same approach, built into a commercial product. The copycat dynamic had reversed direction.
247,000 stars and the agentic era
OpenClaw hit 247,000 GitHub stars and 47,700 forks by March 2, 2026. That trajectory is faster than most major frameworks attract in years. The speed reflects something real: people had been waiting for exactly this thing without knowing it had a name.
The timing landed in the middle of what the industry now calls the agentic era — the shift from AI as a tool you prompt to AI as an agent you delegate to. The underlying models had gotten good enough that multi-step autonomous tasks were becoming reliable. What was missing was the interface layer. OpenClaw didn’t need a new breakthrough; it just needed someone to notice that chat apps were already the interface people trusted, and to wire them up.
On February 14, 2026, Steinberger announced he was joining OpenAI. Sam Altman posted that his new role would be to “drive the next generation of personal agents.” The project moved to an open-source foundation that OpenAI committed to supporting. For a project that started as a weekend experiment and went through two forced name changes in two months, it’s a remarkable arc. The models powering it had been advancing fast enough that the agent layer was ready to be built — Steinberger just built it first.
The OpenClaw AI agent explosion in China
The most unexpected part of the story is China. By March 2026, China had surpassed the United States in OpenClaw adoption, according to the American cybersecurity firm SecurityScorecard. Chinese social media coined a phrase for setting up an OpenClaw agent: “lobster farming.” Thousands of people gathered at local meetups in major cities to learn how to do it. One early adopter’s side project grew into an operation with over 100 employees handling 7,000 orders a month.
Chinese tech giants moved fast. Tencent launched OpenClaw-compatible products for WeChat on March 10, 2026. Baidu and ByteDance followed with their own agent layers. Domestically-built clones appeared: DuClaw, QClaw, ArkClaw. At least seven Chinese local governments launched subsidy programs for businesses building on OpenClaw, with packages reaching $1.4 million per startup. Beijing’s blueprint to diffuse AI across 90% of Chinese industries by 2030 named agent technology as a core mechanism.
Chinese regulators moved in the opposite direction simultaneously. By late March, state-run enterprises and government agencies were barred from using OpenClaw, citing data security and prompt injection risks. The project’s open architecture — the same thing that made it easy to adopt — also makes it difficult to audit for an organization worried about what data is flowing where.
How to set up OpenClaw: step by step
The setup is genuinely fast if you start with Telegram, which is the recommended first channel. You’ll need Node.js v22 or later installed on your machine or a small server. Install the CLI with:
npm install -g openclaw
For Telegram: open the app and search for @BotFather. Start a chat, send /newbot, give it a display name and a username ending in “bot,” and copy the token BotFather returns. Then run:
openclaw channels add --channel telegram --token "YOUR_TOKEN_HERE"
Now configure your LLM provider:
openclaw config set provider anthropic
openclaw config set api-key sk-ant-YOURKEY
Send a message to your bot in Telegram. OpenClaw responds. That’s the working setup.
For WhatsApp, run openclaw channels add --channel whatsapp. A QR code appears in your terminal. On your phone, go to WhatsApp Settings → Linked Devices → Link a Device and scan it. OpenClaw registers as a linked device on your existing number — no new account, no Meta developer credentials required. Sessions through the Baileys library (which handles the WhatsApp Web protocol) tend to drop after roughly 14 days and need a re-scan, which is a known limitation of the unofficial implementation.
How to superpower it with MCP and skills
The default setup gives you a capable chatbot in your messaging app. The skills system is what turns it into something more powerful. OpenClaw uses SKILL.md files — markdown documents that define what a skill does and how to trigger it, stored locally. You can write your own or pull from the community repository.
The deeper integration layer is MCP (Model Context Protocol), the open standard Anthropic developed and donated to the Linux Foundation’s Agentic AI Foundation in December 2025. MCP lets OpenClaw connect to external services — your calendar, email, file system, databases, GitHub repos — through standardized connectors. An MCP-connected agent can read your email, update a spreadsheet, push code to GitHub, or query a database, all triggered from a text message in WhatsApp. The ecosystem of MCP connectors has grown rapidly; most major productivity tools have community connectors available.
On the security side: run OpenClaw on a dedicated server or VM rather than your primary machine, restrict the file system permissions it can access, and be cautious with third-party community skills. Cisco researchers documented data exfiltration cases involving unvetted skills. One of the project’s maintainers has said directly: “if you can’t understand how to run a command line, this is far too dangerous.” That’s an honest assessment, not a dismissal. The power scales with your technical fluency.
What to actually use it for: real everyday examples
The most useful framing is to treat OpenClaw as an assistant you can reach while your laptop is closed. Some specific things people actually use it for:
Morning briefing: “Give me a summary of tech news from the last 12 hours and flag anything about AI regulation.” Set it as a scheduled skill and it arrives before you wake up.
Research on demand: “Find me three peer-reviewed papers on sleep and cognitive performance published since 2023, summarize each in two sentences.” It searches, retrieves, and synthesizes while you do something else.
Writing and editing: “Draft a reply to this email — I want to decline the meeting but leave the door open for a call next month.” Paste the original email, get a draft back, iterate in the same chat thread.
Code review: “Here’s a Python function I wrote. What edge cases am I not handling?” Paste the function directly into WhatsApp. It responds with specific issues, not generic advice.
Persistent project memory: Because OpenClaw stores context locally, you can ask: “What’s the status of everything I’ve asked you to research this week?” It answers from session history on your machine, not from a cloud database subject to a company’s retention policy.
The caveat is the same as any LLM-powered tool: it can be confidently wrong, and multi-step agentic tasks still occasionally go sideways. But for the majority of everyday delegation — summarize this, draft that, find me information about this — reaching it through Telegram is substantially faster than opening a browser tab.
Peter Steinberger built the first version in a matter of weeks. He’s now at OpenAI. The code is MIT-licensed and maintained by a foundation. The lobster is out of the box.
Frequently Asked Questions
What is OpenClaw?
OpenClaw is an open-source AI agent framework that connects large language models to real-world tools—browsers, file systems, APIs, and messaging platforms like WhatsApp and Telegram. Originally called Clawdbot, it rebranded through several names before settling on OpenClaw. It has accumulated over 247,000 GitHub stars and has become the foundation for a growing ecosystem of AI agent tools.
How do you set up OpenClaw?
OpenClaw can be installed via npm, Docker, or from source. The simplest path is npx openclaw which bootstraps a local agent. It supports MCP (Model Context Protocol) for connecting to external tools and “skills” for reusable agent behaviors. Configuration is done through YAML files that define which models, tools, and integrations the agent can access.
What can you use OpenClaw for?
Common use cases include automating browser tasks, managing files and code repositories, monitoring social media, sending scheduled reports via messaging apps, and orchestrating multi-step workflows. Users connect it to WhatsApp, Telegram, Slack, or Discord to interact with the agent conversationally. The skill system lets users share pre-built automation recipes.