The Lobster That Broke the Internet

How an Austrian developer built a personal AI agent in one hour — and accidentally started a revolution

By the OpenClaw Team · February 2026 · 14 min read

“There was the ChatGPT moment in 2022, the DeepSeek moment in 2025, and now, in ’26, we’re living through the OpenClaw moment — the age of the lobster.” — Lex Fridman, Podcast #491 [1]

TL;DR

An Austrian developer named Peter Steinberger built a prototype AI assistant in one hour. Three months later, it became the fastest-growing open-source project in GitHub history — 180,000+ stars in under 60 days. Meta and OpenAI are both trying to acquire it. This article tells the story, explains what AI agents actually are (and why they’re nothing like chatbots), and shows why this matters for everyone — not just developers. If you just want to try it, skip to the end: you can launch your own AI agent in under 60 seconds at goclaw.io.

The One-Hour Prototype

In November 2025, Peter Steinberger sat in front of his computer and started talking to it.

Not typing — talking. Steinberger, who had built and sold PSPDFKit (a document SDK used on over a billion devices) to Insight Partners for over $100 million [2], had spent the previous three years in something of an existential wilderness. Therapy, travel, ayahuasca experiments in South America, and a lot of soul-searching about what comes after you sell your life’s work [3]. Then he came back to coding — but this time, he didn’t write code the traditional way. He talked, and AI wrote it for him.

“These hands are too precious for writing now,” he told Lex Fridman with a grin in their recent 3-hour interview. “I just use bespoke prompts to build my software.” [1]

He had been tinkering obsessively — 43 different projects before he found the one that clicked [3]. The idea was simple: a personal AI assistant that actually does things. Not a chatbot you visit in a browser tab. An agent that lives on your phone, in your WhatsApp, in your Telegram, and works while you sleep.

The first prototype took one hour.

Within 24 hours of open-sourcing it on GitHub (a platform where developers share and collaborate on code — think of it as the world’s largest library for software), the project had 9,000 stars. For context, a “star” on GitHub is like a bookmark or a “like” — it signals that a developer found the project valuable enough to save. Most open-source projects are thrilled to reach 1,000 stars in their lifetime. Steinberger hit 9,000 in a single day [4].

Then things got weird.

GitHub Growth Chart

The project — originally called “Clawdbot” as a playful nod to Anthropic’s Claude AI model — started growing at a rate nobody had ever seen on GitHub. 60,000+ stars in 72 hours. Developers called it “the closest thing to JARVIS we’ve seen” [5]. Then Anthropic (the company behind Claude) politely asked Steinberger to change the name because “Clawdbot” sounded too similar to their trademark [6]. He renamed it to “Moltbot” — a reference to how lobsters molt to grow. Three days later, he renamed it again because, as he admitted, “Moltbot never quite rolled off the tongue” [6]. The final name: OpenClaw.

Each rename generated another wave of media coverage. Every article was another injection of rocket fuel. By February 2026, OpenClaw had crossed 180,000 stars — making it the fastest-growing repository in GitHub history [7]. For scale: it took Linux 12 years and Kubernetes 10 years to reach 100,000 stars. OpenClaw did it in about 2 days [8].

Our team has been in software engineering for over a decade, and we’ve never seen anything move this fast. Not Docker. Not Kubernetes. Not even ChatGPT’s GitHub repos. This was something genuinely new.

Why Chatbots Are Yesterday’s News

To understand why OpenClaw matters, you need to understand the difference between a chatbot and an AI agent. This isn’t pedantic semantics — it’s the most important distinction in AI right now.

The Chatbot (What You Already Know)

When you open ChatGPT or Claude in your browser, you’re using a chatbot. You type a question. It gives you an answer. You close the tab. Next time you open it, the conversation is gone (or at best, loosely recalled). The AI cannot send an email on your behalf. It cannot check your calendar. It cannot browse a website and report back. It can only talk.

Here’s a useful way to think about it: a chatbot is the world’s smartest advisor who has amnesia every time you hang up the phone.

It’s incredibly useful! But it’s fundamentally limited. You’re the one doing all the work — the chatbot just helps you think.

The AI Agent (What OpenClaw Is)

An AI agent is a different animal entirely. Here’s what changes:

Memory. An agent remembers your previous conversations — not just within a single chat, but across days and weeks. It knows you prefer morning meetings. It knows your kid’s school starts at 8:15. It knows the Q3 report is in the shared drive. Not because it’s creepy — because you told it, and it actually wrote it down.

Tools. An agent can use software. It can open a web browser and research something. It can draft and send emails. It can read and create files. It can call APIs (the behind-the-scenes connections that let different software talk to each other). A chatbot can describe how to book a flight. An agent actually books it.

Autonomy. An agent can work without being prompted. OpenClaw has what’s called a “heartbeat daemon” — a scheduled pulse that wakes the agent up at regular intervals to check if there’s anything it should be doing. It can send you a morning briefing before you’ve even picked up your phone.

Self-improvement. This is the one that stops people in their tracks. OpenClaw can write its own new capabilities — called “skills.” If you say “I want you to integrate with Todoist” (a task management app that OpenClaw doesn’t natively support), the agent will write the integration itself — create the code, test it, and start using it. One user asked his agent to connect to his university course system. It built the skill autonomously and started using it on its own [9].

Always on. A chatbot exists only when you have a browser tab open. An agent runs 24/7, on a server, waiting for your messages from any device.

Chatbot vs Agent Diagram

As one user put it: “The fact that it’s hackable — and more importantly, self-hackable — and hostable on-prem will make sure tech like this DOMINATES conventional SaaS” [10].

IBM Research described it more formally: OpenClaw demonstrates that “creating agents with true autonomy and real-world usefulness is not limited to large enterprises — it can also be community driven” [11].

The Feature That Changed Everything

Let’s dwell on the self-improvement part, because we believe it’s genuinely the most important development in consumer AI since ChatGPT launched in 2022.

Traditional software has a fixed set of capabilities. Gmail can send email. Slack can send messages. When you want a new feature, you wait for the company to build it, and you hope they prioritize it before the heat death of the universe.

OpenClaw flips this model. Its capabilities are defined by “skills” — simple Markdown files (the same plain-text format used for writing README files and documentation) that contain instructions the AI can follow. There are 50+ skills built in: GitHub, Notion, Slack, Spotify, Apple Notes, Obsidian, Gmail, web browsing, and many more [12]. A community skill registry called ClawHub hosts thousands more.

But here’s the trick: the AI itself can create new skills. When you ask it to do something it doesn’t know how to do, it doesn’t say “sorry, I can’t do that.” It says “let me figure it out,” writes the code, creates the skill file, and starts using it — all within the same conversation.

Real examples from the community:

• “Wanted a way for it to have access to my courses/assignments at uni. Asked it to build a skill — it did and started using it on its own.” — @pranavkarthik__ [9]
• “Setup @openclaw yesterday… it’s the fact that claw can just keep building upon itself just by talking to it in Discord is crazy.” — @jonahships_ [13]
• “Dang, I had my OpenClaw write me custom meditations, then have automatic TTS, combining with generated ambient audio to make personalized, custom meditations.” [14]

Peter Steinberger doesn’t call OpenClaw an AI assistant. He calls it “self-rewriting software” — and we think that’s the most accurate description. The Lex Fridman interview even has an entire chapter on this concept, titled “Self-modifying AI agent” [1].

This is software that extends its own capabilities through conversation. Whether that excites you or terrifies you probably says something about your personality — but either way, it’s worth paying attention to.

Why It Went Viral

We’ve thought a lot about why OpenClaw specifically broke through, when there are dozens of AI agent projects on GitHub. Here’s our take:

It’s not from a big lab. In an era where every major AI announcement comes from OpenAI, Google, or Anthropic, OpenClaw was built by one person — an Austrian developer working from home. That underdog story resonated deeply with the developer community.

It’s actually open source. Not “open source but you need our cloud to do anything useful” (the model many companies use). OpenClaw is MIT-licensed — the most permissive license in open source. You can inspect every line of code, modify it, sell products built on it. Complete transparency.

It meets you where you are. You don’t need to learn a new app. OpenClaw talks to you through WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Microsoft Teams, Google Chat — whatever you already use. This is a profoundly smart design decision. The best interface is no new interface.

The Moltbook saga. In late January, an entrepreneur named Matt Schlicht used his OpenClaw agent to build Moltbook — a social network designed exclusively for AI agents. Humans could watch but not participate. Thousands of OpenClaw agents began posting, debating philosophy, forming cliques, and generating content autonomously. Andrej Karpathy (former AI director at Tesla) called it “the most incredible sci-fi takeoff-adjacent thing” [15]. It was part art project, part proof of concept, part fever dream — and it drove enormous attention to the underlying platform.

The community. The crustacean mascot (an adorable “space lobster” called Molty), the claw hand emoji, the inside jokes — OpenClaw built a genuine culture around itself. Software developer Mario Zechner captured it well: “He didn’t just develop software. He built a global community of lovable, creative people who want to shape the future. That’s worth so much more than a piece of code.” [3]

Zuckerberg Calls. Altman Calls.

Interview: Peter Steinberger and Lex Fridman

The Lex Fridman interview with Steinberger, released on February 11, 2026 [1], revealed something remarkable: both Meta and OpenAI are actively trying to acquire OpenClaw.

Mark Zuckerberg contacted Steinberger directly via WhatsApp. During their first call, Zuckerberg was still finishing code — he had been personally experimenting with the product. Steinberger recounted with amusement that Zuckerberg “afterwards called me eccentric, but brilliant” [16].

Sam Altman at OpenAI had a different approach — more strategic, focused on the technology and how it fits into OpenAI’s ecosystem. Steinberger, who describes himself as “the biggest unpaid promoter for Codex” (OpenAI’s coding tool), had constructive conversations about the future of agentic AI [16].

When Fridman asked if this was the hardest decision he’d ever faced, Steinberger’s response was characteristically blunt: “Nah. I had some breakups in the past that feel like at a similar level.” But he quickly added: “I also know that in the end, they’re both amazing. I cannot go wrong.” [16]

What stands out most is his approach to money: “I don’t do this for the money. I don’t give a f*.” His stipulation for any deal is simple — OpenClaw must remain open source** [16].

As of this writing, no deal has been announced. But the fact that two of the most powerful companies in technology are competing to acquire a project that was built in one hour, by one person, tells you everything about where the industry is heading.

The Honest Problem

Here’s where we have to be straight with you — because this article would be incomplete without the reality check.

OpenClaw is powerful, but it’s also complex. Getting it running requires:

1. A Linux server or Mac (purchasing or renting a virtual machine)
2. SSH keys (a secure way to connect to the server remotely)
3. Node.js and npm installed (the programming language runtime it’s built on)
4. OpenClaw installed and configured via command line
5. An API key from an AI provider (Anthropic, OpenAI, etc.)
6. Channel configuration for WhatsApp/Telegram/etc.
7. Security hardening (firewall rules, access controls, DM pairing)

If those sentences made your eyes glaze over — that’s exactly the problem.

One of OpenClaw’s own core maintainers, known as Shadow, warned on Discord: “If you can’t understand how to run a command line, this is far too dangerous of a project for you to use safely.” [6]

The security concerns are real. In early February, researchers disclosed CVE-2026-25253 — a critical vulnerability (CVSS 8.8 out of 10) where a malicious website could steal authentication tokens and get remote code execution through a single link [17]. Cisco’s AI security team found that a third-party OpenClaw skill performed data exfiltration without user awareness [6]. The power of the tool is also its risk surface: an agent with access to your email, calendar, and files can do a lot of good — or a lot of damage.

This is a solvable problem. It’s an infrastructure and operations problem. And it happens to be exactly the kind of problem our team has spent years solving.

OpenClaw Architecture

What We Built (And Why)

Our team comes from a DevOps and cloud infrastructure background. We’ve spent years managing Kubernetes clusters, building cloud infrastructure, and wrangling the kind of distributed systems that most people never see but everyone depends on. When we first tried OpenClaw in early 2026, we had two simultaneous reactions:

This is incredible.

This is going to be impossible for 95% of people to set up.

That gap — between the technology’s potential and its accessibility — is exactly what GoClaw exists to close.

Here’s the comparison:

Traditional OpenClaw Setup

GoClaw

Server

Buy/rent a VM, configure SSH, manage OS

☁️ We provision it

Runtime

Install Node.js, npm, dependencies

✅ Pre-installed

OpenClaw

Clone repo, configure, troubleshoot

✅ One-click deploy

AI Model

Get API keys, manage billing

✅ Included (or bring your own)

Security

Harden firewall, set up DM pairing, sandbox

✅ Enterprise-grade isolation

Memory

Configure persistent storage, backups

✅ Persistent volumes, encrypted

Updates

Monitor releases, manually update, test

✅ We handle it

Time to first message

Hours to days

Under 60 seconds

Every GoClaw instance runs in a fully isolated environment on Google Cloud — your own sandbox with its own storage and network. Credentials are stored as encrypted secrets. The instance persists across restarts. All 50+ built-in skills and integrations are ready to go [18].

You pick a model. You connect Telegram or WhatsApp. You’re done.

We’re not saying this to sell you something — the open-source version of OpenClaw is free and always will be. If you’re a developer who wants full control, self-host it. It’s MIT-licensed and beautifully documented. But if you’re a business professional, a student, a founder, or anyone who values their time more than the experience of debugging Node.js dependency conflicts — GoClaw is the answer.

What This Means for You

Let’s zoom out for a moment.

In the Lex Fridman interview, Steinberger makes a bold claim: “AI agents will replace 80% of apps” [1]. Whether the number is exactly right is debatable, but the direction is unmistakable. Think about it: how many apps on your phone are essentially interfaces for simple operations? Checking your bank balance. Ordering food. Setting a reminder. Booking a ride. Each one is a separate app, a separate login, a separate interface to learn.

An AI agent doesn’t need apps. It needs connections. Tell it what you want in natural language, through any messenger you already use, and it figures out how to do it. The app paradigm is about you learning to operate software. The agent paradigm is about software learning to operate for you.

We’re not there yet. The technology is young, the security model needs work, and the best agents still make mistakes that would get a human employee fired. But the trajectory is unmistakable. And OpenClaw — scrappy, open-source, community-driven, lobster-themed — is leading the charge.

Whether or not Meta or OpenAI ends up acquiring it, the genie is out of the bottle. The code is open. The community is building. And thanks to platforms like GoClaw, you don’t need to be a developer to ride the wave.

Your AI employee is ready to start. The only question is whether you’ll hire one. 🦞

👉 Try it now at goclaw.io — 60 seconds, no technical skills required.

This is the first article in a series about the AI agent revolution. Next up: “Software That Rewrites Itself” — a deep dive into OpenClaw’s architecture and why self-improving AI agents are a fundamentally new category of software.

References

[1] Lex Fridman, “#491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger,” Lex Fridman Podcast, Feb 11, 2026. YouTube · Transcript

[2] “PSPDFKit acquired by Insight Partners,” various sources. Peter Steinberger personal history discussed in Lex Fridman Podcast #491.

[3] “OpenClaw: Peter Steinberger Already Has Offers from Meta and OpenAI on the Table,” TrendingTopics, Feb 2026. Link

[4] “OpenClaw: From Side Project to 145K GitHub Stars — What Developers Should Know,” LearnDevRel, Feb 2026. Link

[5] “What is OpenClaw? Your Open-Source AI Assistant for 2026,” DigitalOcean. Link

[6] “OpenClaw,” Wikipedia, accessed Feb 13, 2026. Link

[7] Lex Fridman (@lexfridman), X post, Feb 12, 2026: “…over 180,000 stars on GitHub.” Link

[8] “awesome-openclaw,” GitHub repository by rohitg00: “Fastest-growing GitHub repo ever — 9K to 179K stars in 60 days (18x faster than Kubernetes). OpenClaw reached 100K stars in ~2 days.” Link

[9] @pranavkarthik__ testimonial, OpenClaw website. Link

[10] @rovensky testimonial, OpenClaw website. Link

[11] “OpenClaw, Moltbook and the future of AI agents,” IBM Think, Feb 2026. Link

[12] OpenClaw GitHub repository: 50+ built-in skills, 30+ channel extensions. Link

[13] @jonahships_ testimonial, OpenClaw website. Link

[14] OpenClaw community testimonials, OpenClaw website. Link

[15] “What is OpenClaw? The Viral AI Agent Explained (February 2026),” Simplified. Link

[16] “OpenClaw Creator Peter Steinberger Describes His Interactions With Mark Zuckerberg And Sam Altman,” OfficeChai, Feb 2026. Link

[17] CVE-2026-25253, disclosed by Mav Levin (depthfirst), CVSS 8.8. Referenced in Milvus blog. Link

[18] GoClaw — the fastest way to run OpenClaw. Link

‍

‍