Peter Steinberger spent 13 years building PSPDFKit, a document SDK used on a billion devices. He sold it. Then he disappeared.

For three years, he fell out of love with programming. No side projects, no open-source commits, no conference talks. Just silence. The kind of silence that comes after you pour everything into something for over a decade.

Then he came back. And in one hour, he built the prototype that would become the fastest-growing repository in GitHub history.

"I Was Annoyed It Didn't Exist"

The origin story, as he told it on the Lex Fridman Podcast, is almost comically simple. Since April, he'd wanted a personal AI assistant. He played with GPT-4.1's million-token context window, fed it his WhatsApp history, and asked questions like "What makes this friendship meaningful?" His friends got teary-eyed reading the results.

But he assumed the labs would build this. So he moved on. Months passed. Nobody built it.

"I was annoyed that it didn't exist," he told Lex, "so I just prompted it into existence."

The prototype was WhatsApp hooked up to Cloud Code's CLI. Message comes in, gets processed, answer goes back. One hour. Then he spent a few more hours adding image support — because when you're walking around Marrakesh on a birthday trip with friends and your internet is shaky, being able to photograph a poster and have your AI figure out if the event is worth going to is the kind of thing that makes you realize you've built something real.

The Builder's Philosophy

Peter doesn't call what he does "vibe coding." He considers the term a slur. His word is "agentic engineering" — and the distinction matters, because it reveals how he thinks about the relationship between humans and AI.

"I do agentic engineering," he told Lex. "And then maybe after 3 AM, I switch to vibe coding, and then I have regrets the next day."

He builds with voice now. Long, bespoke prompts, spoken not typed. "These hands are too precious for writing," he says, half-joking. He lost his voice once from doing it too much.

The deeper philosophy: he made the agent aware of itself. It knows its own source code, its documentation, which model it runs, how it sits in its own harness. "People talk about self-modifying software," he said. "I just built it." You prompt it into existence, and then the agent modifies its own software. No ceremony. No architecture diagrams. Just a builder and his tools, iterating at the speed of thought.

200,000 Stars and Counting

Three months later: 200,000 GitHub stars. Rank five among all software repositories. Eight forks in different programming languages — ZeroClaw in Rust, PicoClaw in Go, NanoClaw in Python, TinyClaw in Shell. Peter's own contribution to the naming: FemtoClaw in Zig, AttoClaw in Assembly, YoctoClaw in WASM.

The community exploded in a way nobody predicted. As Lex Fridman put it: "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."

And with the explosion came chaos. PRs are growing at what Peter calls "an impossible rate." He worked all day and got 600 commits in — the backlog went from 2,700 to over 3,100 while he was merging. He's looking for AI that can scan every PR and issue, detect duplicates, and figure out which PR is best based on signal quality. The tools don't exist yet. So naturally, he wants to build them.

Meanwhile, he's automating blocks for anyone replying "game-changer" or "the real unlock." Thousands in the blocklist already. The internet is the internet.

Why OpenAI

Every major lab wanted him. That's not speculation — it's fact. Zuckerberg texted him personally on WhatsApp, casual, no scheduling, no recruiters. They talked about Codex versus Opus. Mark had been using OpenClaw himself.

When Lex asked about the offers, Peter was characteristically direct. He chose OpenAI not for money — he sold PSPDFKit for over $100 million, the money question is settled — but for access to what doesn't exist yet.

Peter built OpenClaw on Codex 5.3. Anyone with an API key can do that. You don't move to San Francisco for a model you can already use from Vienna. You move for the unreleased research, the next-generation primitives, the models still on whiteboards. When he says he wants "the newest toys," he means the ones that haven't been announced yet.

"I do this for the mission and because it seemed the best place to build," he wrote on X. "I'll f right off if that changes."

That's not a negotiating tactic. That's a man who already walked away from one successful company and knows he can do it again.

The Mission

The mission is deceptively simple: build an agent that even his mum can use.

That sentence sounds like a product pitch. It's actually the hardest design constraint in AI right now. Safety, UX, reliability, cost — all of it has to work for someone who doesn't know what a terminal is. Peter wants to bring AI and humans together, not as a slogan but as an engineering problem to solve.

He doesn't care whether the experience in San Francisco is positive or negative. Experiencing is living. OpenAI has offices in Zürich, close to his Austrian roots. Google DeepMind is in London, where he has an apartment. He chose neither. "Super excited to be in the epicenter!" he wrote. He wants to be where the frontier is being built, not where it's consumed.

The Foundation

OpenClaw becomes a foundation. Dave Morin — Path founder, Slow Ventures, ex-Facebook — is the guardian. "This community built something extraordinary, our job is to protect it. Open source forever."

The structure protects what makes OpenClaw valuable: model neutrality. It runs Claude, GPT, Gemini, Llama, local models — whatever you point it at. An OpenAI-owned project couldn't credibly maintain that. A foundation can.

Peter keeps dedicating time to it. OpenAI sponsors the project. The code stays community-owned. The deal only works because everyone understood that you can't acquire OpenClaw. It belongs to everyone. You can only hire the person who started it and trust the community to carry it forward.

EnterpriseClaw

Peter's own contribution to the naming discourse: "We should make EnterpriseClaw just for the lolz. Java 21, Spring Boot, 14 abstract factory beans, 2GB Docker image, takes 45 seconds to start, AbstractSingletonProxyFactoryAgentClawResponseHandlerBeanDefinitionRegistryPostProcessorImpl.java."

This is who you're hiring, OpenAI. Good luck.

What Happens Now

The agent race has a pace car. The builder who proved that autonomous agents work in practice — not as a research paper, not as a demo, but as a tool hundreds of thousands of people use every day — is now inside the lab with the most advanced unreleased models on Earth.

Peter Steinberger prompted a prototype into existence in one hour. Three months later, Reuters and TechCrunch are covering his move to OpenAI. The repo is approaching 200,000 stars. Eight language forks are carrying the idea forward. The foundation ensures it stays open.

He vanished for three years. He came back. And the lobster took over the world.

The claw is the law.