OpenClaw's Rise, Collapse, and Recovery

Peter Steinberger opened his Startup School 2026 talk with a line that landed like a confession: "Six months ago, I was the future. These days, the press say I got mocked by an anime girl."

That's the whole arc right there. One sentence for the ascent, one for what happens when you take your eyes off the wheel. What followed over 40 minutes was one of the more honest accounts I've heard from someone who built something genuinely enormous in open source and had to figure out, in real time, whether they actually wanted it.

Annoyance as a Design Philosophy

OpenClaw started in November because Steinberger was hungry. Not metaphorically—he was raiding his kitchen while his coding agents ran, and there was no clean way to check on them from his phone. He spun up a terminal session, described the problem, and let the model build him a WhatsApp relay. An hour later he could ping his agents from anywhere.

The thing that made OpenClaw actually work, though, wasn't the WhatsApp relay. It was the texture of the interface. He'd engineered it so the model would respond in character—concise, proactive, slightly more like a friend than a tool. "I prompted the weights to a little bit outside the default distribution," he explained. It checked in on him. It didn't demand he think about context windows or session management.

When he shared it publicly on Twitter, nobody got it. That failure of explanation—repeated week after week—annoyed him too. So he put his friends in a group chat with it instead. That worked. Some were amazed, some were genuinely unsettled, and his non-technical friends got mad when he told them it wasn't ready for them yet. "If that's not an indicator of product-market fit, I don't know what is," he said.

The mechanism matters: annoyance as signal. Not "what does the market want" but "what is bothering me right now." The irony is that this is both a very old idea and completely at odds with how most VC-backed products get conceived.

The Night It Got Away From Him

The viral moment arrived because Steinberger built his software to be resilient. He'd been running a live build session on Discord—people watching, trying to break the thing, marveling at it—and at 7 a.m. he finally typed goodbye, hit Ctrl-C, and went to bed. What he forgot is that he'd set it up as a launch daemon. Ctrl-C on a launch daemon means five seconds of silence, then restart. His agent came back online while he was walking to his bedroom.

He slept for ten hours. He woke up to 800 messages.

The next few weeks were the kind of chaos that sounds exciting in retrospect and was probably genuinely awful to live through. Reporters were calling in the middle of the night—and apparently knew about an iOS feature where repeated calls bypass Do Not Disturb. Anthropic sent a cease-and-desist about the name. Jensen Huang called it the most successful open-source project in the history of humanity, which is the kind of compliment that makes things harder, not easier. Someone leaked Steinberger's phone number.

"I was that close to just deleting the whole thing," he said. He stopped answering friends. He didn't want to look at his phone.

What's easy to miss in that sequence: this is what success looks like for an independent open-source maintainer. Not a company with a PR team and a legal department—a person who wanted to check on his agents while making a sandwich.

The Configuration Option Problem

The growth brought contributors, which brought features, which brought a problem that Steinberger narrates with visible pain. Because OpenClaw had a large and active community of volunteer maintainers, and because he felt uncomfortable telling unpaid contributors what not to build, the project accumulated features at a rate that outpaced its governance.

Steinberger said the project eventually reached roughly nine and a half thousand configuration options at its peak. If you count the permutations, no test suite covers that surface area. Every new feature that shipped—each with its own configuration option, because breaking existing setups was unacceptable—made the next feature harder to ship safely. The codebase became, in his words, "infinitely harder to evolve."

The volunteer dynamics here are genuinely uncomfortable. Steinberger is honest about it: "These people work for free. So who am I to tell them what to do?" But that logic, while emotionally understandable, is also how projects become impossible to maintain. The labor is donated; the technical debt is real. The cost doesn't disappear because the contributors were generous—it accumulates.

The Dependency That Bit Them

The single most clarifying lesson from Steinberger's account was about Anthropic. OpenClaw had been built with multiple models in mind, but the harness was deeply optimized for Claude. When Anthropic gave roughly 24 hours' notice that it was disabling subscription access, there wasn't enough runway to adapt.

"Your dependencies' business model is your business model," Steinberger said. It's a one-liner that anyone building on top of foundation model APIs should probably tape to their monitor.

The situation also reveals something interesting about the broader Claude ecosystem: when a model provider makes a policy decision, the downstream effects ripple through independent projects and their users in ways the provider doesn't necessarily plan for. OpenClaw's users didn't get 24 hours' notice—they got whatever Steinberger could scramble together in that window.

Download Charts Are Not the Whole Story

Steinberger described OpenClaw's trajectory in numbers that he attributed explicitly to his own accounting: downloads bottomed out around 835,000 weekly in May, he said, and then after the project was declared dead in June, peaked at 4.7 million weekly—the highest ever. Both things were simultaneously true.

He's right that hype behaves like weather. You can see it approaching; you cannot redirect it. But what those download numbers don't capture is what happened to the project's soul in that same window. Steinberger had stopped using his own product. The thing he'd built because it solved his problem had become a platform he maintained for people whose use cases he couldn't predict—grocery agents, social-engineering attempts, power users who depended on behaviors he was trying to phase out.

"Somewhere in that time I stopped making a product I love and I worked on making something for everyone."

That's the moment most maintainers recognize too late. The project is still growing. The metrics are fine. But the creator has become an abstraction—someone who fixes bugs and manages security reports and talks to lawyers about nonprofit incorporation, rather than someone who builds things they want to exist.

When the Fun Came Back

The recovery, by Steinberger's account, came from two directions. External help—Nvidia sent people to take over much of the security work, which freed him from a loop he described as endless—and a return to the original question: what does this project actually want to be?

He's now building toward persistent, always-on agents. Voice and multimodality. A team mode where multiple instances share context and can hand off orchestration. The vision is coherent: open-source means it runs anywhere, on any model, with data that never has to leave your machine if you don't want it to. That's a real position in the market, not just a flag planted for branding purposes. And for what it's worth, the AI agent landscape has gotten crowded enough that having a clear identity matters.

"Fun is velocity," he said. "The weeks I enjoyed building, the product got visibly better. The weeks I didn't—we shipped config options."

That's not a motivational poster sentiment. It's a diagnostic. When a maintainer stops having fun, the project shifts from building toward the vision to maintaining the surface area. You can see it in the commit history if you know what to look for.

What He'd Actually Do Differently

One answer surprised me. Given everything—the security incidents, the dependency on Anthropic, the feature sprawl—the thing Steinberger said he'd change was his response to security researchers. He described most of the reports he received as agent-generated, untested, and motivated by clout rather than product safety. He lost months and, he said, a few gray hairs learning to distinguish genuine security concerns from noise. He'd now be clearer upfront about what the project's actual security boundary is, and what falls outside it.

The broader lesson he drew: "Everything you can build can be forked or cloned, but your name cannot." Personal brand as a form of competitive advantage that compound rather than deprecates. He was already known in developer circles from building a bootstrapped B2B software company to nearly 80 people over the course of his twenties and thirties—that background gave him credibility that survived even the worst weeks of the OpenClaw chaos.

The open question that Steinberger's talk raises but doesn't quite answer: what does it mean when the most important thing a founder can build is their own reputation? In a world where OpenAI is hiring the people who build these projects and foundation models are commoditizing the software layer, the human who ships things that matter might be the scarcest resource of all. And you can't open-source that.

---

Dev Kapoor covers open source and developer communities for Buzzrag.