How Peter Steinberger Built OpenClaw Into the Largest AI Skill Registry

In February 2026, Peter Steinberger announced he was joining OpenAI and transferring OpenClaw to an independent foundation. The announcement described a "playground project" that had grown into something he hadn't anticipated.

By then, OpenClaw had 220,000+ GitHub stars, 13,000+ skills in its registry, and 1.5 million downloads. Multiple companies — including one of the largest in the world — had built products on top of his code. Sam Altman had called him a genius. He'd received offers from OpenAI, Meta, and Anthropic simultaneously.

That's not how most playground projects end.

The Long Background

Peter Steinberger studied medical computer science at TU Wien — Vienna University of Technology. He stayed in Vienna after graduation, started writing iOS software, and in 2011, while waiting for a U.S. work visa, built PSPDFKit as a weekend project.

PSPDFKit became a serious business. The SDK powered document processing for Dropbox, IBM, and a long list of enterprise clients, handling billions of document interactions annually. Steinberger ran the company for 13 years. In 2023-24, he sold it for approximately €100 million.

The exit should have felt like success. By most measures it was. But the transition from founder to post-exit didn't go smoothly.

In a Fortune profile published after the OpenAI announcement, Steinberger described the period after selling PSPDFKit in terms that resonated with anyone who has experienced founder burnout. He'd spent over a decade building one thing, and when it was gone, the engine that had driven him stopped. The comparison he reached for — Austin Powers losing his mojo — captures something real about what happens when identity and work become too tightly fused.

The Return: From AI Experiments to OpenClaw

Steinberger's blog, steipete.me, serves as a public record of his return to building. Posts from late 2025 track the evolution from individual AI experiments to systematic agentic workflows to OpenClaw itself.

By October 2025, he was writing about what he called "just talking to it" — his philosophy for working with AI models as creative collaborators rather than autocomplete tools. By December 2025, he'd published a detailed account of how he restructured his development workflow around inference speed rather than personal coding speed. He was running 3-8 projects simultaneously by queuing tasks across AI agents. He was no longer reading most of the code his agents wrote; he was watching the output stream and intervening at key decision points.

The key insight he articulated in December: inference speed now limits productivity more than individual coding ability. The developer's critical skill had shifted from writing code to choosing architectures, selecting frameworks, and making decisions that agents could execute well.

This thinking directly shaped OpenClaw's design.

Building OpenClaw

The project that became OpenClaw started in late 2025 under a different name. The original name, Clawd, prompted a trademark dispute with Anthropic. The community renamed it Moltbot — a reference to the lobster lifecycle of molting and growing. That name lacked resonance. The final name, OpenClaw, honored the lobster mascot, cleared trademark review, and stuck.

Steinberger averaged 6,600 commits per month during peak OpenClaw development. He managed a 300,000+ line TypeScript codebase with AI assistance, applying the same agentic workflows he'd written about publicly. The project was built by the same methods it was designed to enable.

The SKILL.md format — the core innovation that made ClawHub possible — reflects his philosophy of radical simplicity. A skill is a directory with a markdown file. That's all. Any developer who can write markdown can write a skill. Any AI that can read markdown can learn from a skill. The barrier to contribution is a few minutes, not a few days.

The skill creation flywheel emerged naturally from that design choice. Users asked for capabilities. OpenClaw generated SKILL.md files through conversation. Those files got published. The registry grew.

The Problem of Success

By early 2026, OpenClaw had 145,000+ GitHub stars and 2 million visitors in a single week. The project had grown well beyond what one person could steward.

Steinberger was also paying approximately $10,000 per month in server costs. He described the project's scale as something that surprised him — the gap between "playground project" and "infrastructure that millions of people depend on" had closed faster than he'd expected.

The offers came from multiple directions. OpenAI, Meta, and Anthropic all approached him. He chose OpenAI, explaining that joining the company working on the fastest path to general AI capability was the quickest way to achieve his stated mission: building an agent that anyone — including his mother — could use.

The Foundation Model

The transfer of OpenClaw to an independent foundation, sponsored by OpenAI, is worth examining separately from Steinberger's personal decision.

The foundation model keeps OpenClaw open source while providing governance that doesn't depend on a single person's availability or motivation. OpenAI's sponsorship covers infrastructure costs that had been running at five figures per month. The OpenClaw community — which had contributed 5,400+ skills and significant code — gets an institutional home that doesn't belong to any single company.

This is the mature move for an open-source project of OpenClaw's scale. The Linux Foundation, Apache Foundation, and similar structures exist precisely because valuable open-source infrastructure needs governance that outlasts individual creators.

Steinberger's choice to hand OpenClaw to a foundation rather than sell it to a single acquirer or let it run as a personal project preserved the ecosystem's independence. That independence matters for the 13,000+ skill creators who built on top of his work.

What He Built That Matters Most

The GitHub star count and the Fortune profile are the visible outputs. The more durable contribution is the SKILL.md format and the ClawHub registry.

Steinberger gave the AI agent ecosystem a standard for packaging and distributing behaviors. Before SKILL.md, extending an AI agent meant custom integrations, proprietary plugin formats, or complex API configurations. After SKILL.md, it means writing markdown.

The registry that grew around that format reached 13,729 entries in six months. The community that built those entries represents the collective intelligence of thousands of developers figuring out what AI agents should be able to do.

Tencent noticed. OpenAI noticed. The 220,000 people who starred the repository noticed.

Peter Steinberger built a playground project and accidentally defined an industry standard. That's a career highlight that even a €100 million exit can't match.