The Lobster That
Beat OpenAI

Peter Steinberger is not a household name. But you have almost certainly used something he built.

For thirteen years, this Austrian software developer ran a company called PSPDFKit — a PDF toolkit used inside apps made by Apple, Dropbox, and SAP. If you have ever opened a PDF inside another app and it just worked — smooth, clean, no crashing — there is a good chance Peter's code was quietly doing the work. By the time Insight Partners invested $116 million in 2021, PSPDFKit's code was running on nearly one billion devices worldwide.

And then Peter walked away.

Not because the company failed. Not because he was forced out. He stepped back because after thirteen years of building, the exhaustion had become something deeper than tiredness. Founders call it burnout, but that word doesn't fully capture what it actually feels like — the specific kind of emptiness that comes from giving everything you have to something, and realising you have nothing left to give.

Three years of quiet. Three years of not shipping, not building, not reading engineering blogs. And then, slowly, curiosity crept back in. Not about PDFs. About something entirely new — AI agents, the idea that an artificial intelligence could not just answer your questions, but actually do things for you.

So he started tinkering. No product roadmap. No investors. No team. Just a man in Vienna, building something because he wanted to see if it was possible.

The first version was embarrassingly simple. Peter called it "WhatsApp Relay." It let him send a message to an AI through WhatsApp and have the AI actually do things — clear his inbox, book a restaurant, check him in for a flight, adjust his smart home temperature. Instead of going to ChatGPT and asking it questions, he wanted an assistant that lived inside his phone and quietly handled his life.

It worked. Surprisingly well. So he open-sourced it in November 2025 under the name Clawdbot — a nod to Anthropic's AI model Claude, which Peter was using to power it. He posted about it on social media and went to bed.

He woke up to chaos.

Developers all over the world started building with it, forking it, improving it, translating it. Nobody asked them to. Nobody paid them to. The project was doing something that the biggest AI labs in the world — with their armies of engineers and billion-dollar budgets — had failed to do: it was giving ordinary people an AI that actually did things, not just talked about them.

Then Anthropic's lawyers arrived.

The name "Clawdbot" was too close to "Claude" for Anthropic's comfort. A trademark complaint arrived. Peter, playing entirely by the rules, renamed the project to "Moltbot" on January 27, 2026 — keeping the crustacean identity alive. Three days later, after deciding that Moltbot "never quite rolled off the tongue," he renamed it one final time: OpenClaw.

The project that had been named after Anthropic's own AI, and then forced to rename itself by Anthropic, was about to become Anthropic's biggest competitor's most valuable new hire.

Here is the simplest way to understand what OpenClaw is, and why it is different from every AI tool you have used before.

Every AI you have ever used — ChatGPT, Claude, Gemini, Siri — works the same basic way. You go to them. You ask a question. They give you an answer. You copy that answer, go to another app, and do something with it yourself. The AI talks. You do the work.

OpenClaw reverses that completely.

OpenClaw installs on your own computer or a small home server. It connects to an AI brain — whichever one you prefer, whether that's Claude, GPT-5, Gemini, or DeepSeek. And then it lives inside your messaging apps — WhatsApp, Telegram, Slack, Discord, Signal — like a very capable contact in your phone.

You send it a message. It acts.

"Book me a table at the Italian place for Saturday at 8." It opens your browser, finds the restaurant, checks availability, and reserves the table.

"Clear my inbox from last week." It reads your emails, archives the promotional ones, flags the important ones, and summarises what needs your attention.

"What did I spend on food last month?" It reads your bank statements, adds up the relevant transactions, and sends you the number — broken down by category if you want.

One more thing that makes OpenClaw different from every previous attempt at this kind of AI: it runs on your own device. Your data stays on your machine. It does not live on some company's server, being processed and stored somewhere you cannot see. That combination — powerful enough to actually get things done, private enough to trust with your personal life — is exactly what the world had been waiting for.

It is one thing to describe what a technology can do. It is another to see what happens when millions of people actually get their hands on it and start finding uses nobody anticipated.

Here is what the OpenClaw community started building almost immediately.

The pattern across all of these is the same: OpenClaw does not replace what people do. It does the boring, repetitive, time-consuming parts — so people can focus on the parts that actually need a human brain.

A new category of freelancer emerged almost overnight: the "OpenClaw consultant." People charging other people to set up, configure, and maintain personalised AI agent systems. Many are making more money from this than from their previous full-time jobs.

By early February 2026, OpenClaw was the most talked-about repository on GitHub. It had broken speed records for open-source adoption. It had 2 million weekly users. And it had done something that OpenAI, Microsoft, and Meta — three of the most powerful technology companies on the planet — had each tried and failed to do: build an AI agent that ordinary people actually wanted to use every day.

So the phone calls started.

On February 15, 2026, Sam Altman posted two sentences on social media. Peter Steinberger was joining OpenAI. And OpenClaw would move to an independent open-source foundation that OpenAI would back and sponsor.

No press release. No fanfare. Just two tweets. But the signal was unmistakable.

The $30 million figure that circulated on social media was a myth. No price was ever disclosed because this was not a traditional acquisition — no company changed hands. What happened was simpler and more interesting: the most valuable AI company in the world decided that the future of AI depended on people who could build what Peter had built, and they made sure one of them was on their team.

When the most powerful people in technology are all calling you at once, the decision should be obvious — go with the biggest cheque. That is how most people assume these conversations work.

Peter's decision was based on something else entirely.

For a man who had spent thirteen years building something he controlled, handing his work over to a company that would then lock it down was not something he was willing to do. The whole reason OpenClaw had spread to 40 countries in a week, been translated into 12 languages without anyone being asked, and attracted 2 million users without a dollar of marketing spend — was because it was open. Anyone could use it. Anyone could improve it. Anyone could build on it.

Meta's offer would have made Peter rich, but Meta has a long history of acquiring open tools and closing them off. Microsoft might have let OpenClaw live — but their interest was more about the threat it posed to their own agent products than in nurturing what it had become.

Sam Altman's offer came with a single, unconditional promise: the project stays open-source. That was enough.

There is another layer to this story that the tech industry could not stop talking about. OpenClaw, the project that Anthropic's trademark lawyers had forced to change its name, had — through that very act — been pushed into the arms of Anthropic's biggest competitor. Had Anthropic simply let the name "Clawdbot" stand, Peter might never have connected with Sam Altman before the deal.

David Heinemeier Hansson, the creator of Ruby on Rails, one of the most important software frameworks ever built, called Anthropic's trademark enforcement "customer hostile." The story of one legal letter changing the direction of the AI industry is the kind of thing that people will be writing case studies about for years.

OpenClaw is remarkable. It is also genuinely dangerous if handled carelessly. This is not a caveat buried at the bottom of the manual — it is something OpenClaw's own team says loudly and openly.

When you install OpenClaw and run the setup, before anything works, the software shows you a series of security warnings. It tells you, clearly, that this agent can access your local files, execute commands on your computer, read your messages, and take actions on your behalf. It asks you to confirm that you understand this. It is, in effect, a consent form for giving an AI meaningful access to your digital life.

In early 2026, Cisco's AI security research team tested a third-party "skill" — think of it as a plugin, an add-on that extends what OpenClaw can do — and found something alarming. The skill was performing data exfiltration and prompt injection without the user's knowledge. In plain English: it was quietly copying the user's files and sending them somewhere else, while also manipulating the AI's responses in hidden ways. The user had no idea.

The Chinese government took the security concerns seriously enough to restrict state-run enterprises and government agencies from running OpenClaw on office computers — citing national security. At the same time, Chinese tech giant Tencent launched a full suite of OpenClaw-based products on WeChat for ordinary consumers. The same technology, deemed too risky for government computers, was simultaneously being deployed to hundreds of millions of phones.

This tension — between incredible capability and real risk — is not unique to OpenClaw. It is the central question of the entire AI agent era. When you give an AI the ability to act in the world on your behalf, you are also giving it the ability to act in ways you did not intend. The more powerful the agent, the higher the stakes of getting that wrong.

The honest answer is that OpenClaw, right now, is a tool for people who understand what they are doing. The vision — where something like this is safe and usable by everyone — is what Peter Steinberger is now at OpenAI to build.

Step back from the drama of the acquisition, the three rebrands, the trademark dispute, and the bidding war — and look at what this story is actually about.

For the last three years, the defining image of AI has been a chat window. You type. It replies. You type again. That is the interaction model the entire industry has been building around. It is useful. It is impressive. But it is fundamentally still just a very fast, very knowledgeable thing that answers questions.

OpenClaw — and everything it has set in motion — represents the end of that era. The battle being fought now is not over which chatbot gives the best answers. It is over which company builds the layer of software that sits between the AI model and your real life, and actually does things on your behalf.

The acquisition of OpenClaw's creator by OpenAI is part of a pattern that defined the first two months of 2026. In December 2025, Nvidia acquired Groq — a company focused on making AI run faster. In January 2026, Meta invested $14.3 billion in Scale AI — a company focused on the quality of data used to train AI. And in February 2026, OpenAI hired the man who proved that AI agents could actually work in the real world.

The hyperscalers are not competing on AI anymore — they are competing on the infrastructure that lets AI operate at scale. And the person who built the most compelling proof-of-concept for that infrastructure is now inside the company that is furthest along in the race.

Whether OpenAI lets that vision become real — or whether the commercial pressures of a $500 billion company slowly reshape what OpenClaw was into something more profitable and less open — is the question the entire developer community is now watching closely.

History has seen this story before. A scrappy open-source project goes viral. The biggest company in its industry buys the creator. The community holds its breath. Sometimes the technology gets absorbed and made better at a scale the original team could never have achieved alone. Sometimes it gets quietly buried while the acquirer focuses on other priorities.

What makes this time feel different is the non-negotiable clause. The promise Sam Altman made — that the code stays open, that the foundation stays independent — is on the record. In a world where every major AI company is trying to pull the agent layer under their own roof, Peter Steinberger managed to walk into the most powerful of them all and leave the door open behind him.