Transcription
The lobster is joining the lab. Here's what OpenAI's hire of Peter Steinberger, inventor of Open Claw, actually means.
So, Peter Steinberger built the fastest growing open-source project in GitHub history in his living room, all while bleeding $20,000 a month. Then, on Valentine's Day, he posted three paragraphs on his personal blog announcing he was joining OpenAI. Sam Alman followed up with a post on X calling Steinberger a quote genius that would drive the next generation of personal agents.
The announcement landed less than 48 hours after OpenClaw shipped its most significant security update ever, patching more than 40 vulnerabilities across the frankly brand new platform. This is not a coincidence, and it is not primarily an aqua hire, even though it's being painted as such. What happened this weekend is the clearest signal yet of where the AI industry is headed in 2026, which is frankly away from chat bots, toward agents that do real work on real computers, and toward the brutal competitive question of who gets to own the platform layer underneath them.
So to understand why this hire matters, you need to understand three things. First, what Open Claw really is at root. Second, why open AAI needed it more than Steinberger needed OpenAI. And third, what changes now for the roughly 200,000 developers who had started the project on GitHub.
First, let's step back to the Friday night hack that produced 200,000 GitHub stars. The origin of Open Claw borders on the absurd. On a Friday night in November of 2025, Peter Steinberger, who's an Austrian developer who had already sold his PDF framework company for over a hundred million, sat down and built a prototype in about an hour. The concept was relatively simple: just wire up a large language model into WhatsApp so he could read messages, browse the web, and execute shell commands on your behalf.
Steinberger had spent three years deliberately away from tech after his exit. He went traveling, he did therapy, he experimented with Iawasa, and he cycled through what he's described as a period of deep searching. He came back to coding because AI pulled him back. Before Open Claw, he'd turned through 43 different projects. Think about that the next time your idea fails. Number 44 turned out to be the one for him.
The initial version was quite crude. It connected a messaging interface to Claude and Thropics language model and it could do basic tasks. Ber open sourced it as quote Claudebot on on Claude and the lobster claw that became the project's mascot. By mid January, it had around 2,000 GitHub stars, which was respectable for a side project, but not yet a phenomenon. Then everything detonated at once.
On January 27, Anthropic's legal team sent a trademark notice. Claudebot sounded too much like Claude. Steinberger agreed to rename. He chose Maltbot. Lobsters malt when they outgrow their shells. During the renaming process, in the literal few seconds between releasing his old GitHub handle and claiming the new one, crypto scammers sniped the account and began promoting a fake Clawude token on Salana. They served malware from his GitHub. They hijacked his npm packages. Steinberger nearly deleted the entire project.
Three days later, he renamed the project one more time to Open Claw. This time with purchase domains, completed trademark searches, and a coordinated account exchange, which was executed with what he described as Manhattan Project level secrecy. He also spent $10,000 to buy a dormant Twitter business account to secure the handle.
All of this chaos, paradoxically, was the accelerant. Each rename triggered new threads on Reddit and Hacker News. The the trademark drama drew even more media coverage. A simultaneous launch of Multibook, an experimental social network designed exclusively for AI agents, went viral with Fortune, CNBC, and TechCrunch, all covering AI bots creating their own religions, their governments, their existential poetry.
Within weeks, Open Claw had soared, crossing a 100,000 GitHub stars, and it just kept going. Now it's up to 200,000 stars and it is easily the fastest growing repository in GitHub's history with more than 10,000 commits from 600 contributors in under three months. The numbers do matter, but they obscure a more important point. OpenClaw proved that a self-hosted AI agent could do things that no chatbot had done before. Ultimately, that's why it grew.
It didn't just answer questions for users. It managed emails, scheduled meetings, controlled browsers, executed shell commands, sent messages across WhatsApp, Telegram, Slack, Discord, Signal, and iMessage. It ran on your hardware if you wanted it to, and it stored your data locally, again, if you wanted it to. And its most unsettling feature, the one that made researchers both excited and alarmed, was that it could modify its own source code.
So, that's the story of Open Claw. Why OpenAI over Meta? Because both OpenAI and Meta were competing for Peter Steinberger's attention and for a stake in that open- source framework. Before dissecting what OpenAI got, it's worth understanding the negotiation Steinberger walked through because the way he chose reveals what this deal is really about.
Both Mark and Sam made concrete offers. Zuckerberg reached out via WhatsApp initially and when Steinberger suggested they just call right then instead of scheduling, Zuck asked for a few minutes, apparently because he needed to finish coding. That detail clearly resonated for Peter, a founder who built his reputation on shipping. Zuckerberg tried OpenClaw personally and sent a message calling it amazing. He also gave really blunt product feedback, alternating between praise and pointed criticism. And Peter valued that directness, noting that hands-on engagement showed that Zach actually cared about the product. I mean, I have yet I have yet to meet a developer that didn't appreciate specific and useful product feedback.
On the OpenAI side, Sam's pitch came with something more tangible. A promise of computational power tied to the Sarah deal that could dramatically accelerate agent performance, plus the fact that OpenAI was already contributing tokens to the project. It got lost in all of the shuffle, but Peter Steinberger built Open Claw using codecs.
Steinberger described his conversations with Alman as thoughtful and substantive, and he also acknowledged having more personal connections at OpenAI and a deeper history of building on their tech. But Peter admitted that he didn't get the same hands-on product engagement from OpenAI that he had gotten from Zuck. The deciding factor appears to have been mission alignment rather than personal chemistry.
Steinberger's stated goal, building an agent his mother could use, requires access to frontier models and the kind of research pipeline only a hyperscaler lab could provide. He spent the week before the announcement in San Francisco meeting with labs getting access to unreleased research. And fundamentally, he came away saying OpenAI's vision most closely matched his own. Critically, OpenAI also agreed to support OpenClaw as an independent open- source project through a foundation, which preserves the condition Steinberger called non-negotiable from the start. The last thing he wanted was to have Open Claw become closed source, to close the claw.
His attitude throughout the process was characteristically blunt. When Freriedman asked him if this was the hardest decision he'd faced, Peter Steinberger replied, "Nah." The man who already had a nine-figure exit and had spent three years soulsearching before returning to code does not apparently agonize over these kinds of career moves. The beauty is if it doesn't work out, I can just do my own thing again, he told Freriedman. That posture, complete optionality, zero desperation gave him more leverage than most Aqua hire candidates ever have.
So, what did OpenAI actually get? Open AAI did not acquire OpenClaw. I want to be precise about that. Steinberger, the individual, is joining the company as an employee. Open Claw itself is moving to an independent foundation and will remain open source. Altman confirmed on X that OpenAI will continue to sponsor the project. This is a really important distinction.
OpenAI got Steinberger his vision, his developer credibility, his community influence, and his proven ability to build agentic systems. people actually use. What they did not get is exclusive control of OpenClaw, the platform. The Chrome and Chromium model that Seinberger had floated in his Lex Freedman interview appears to be roughly what's taking shape. In other words, OpenClaw as the open source foundation, like Chromium, the engine that powers Chrome, and OpenAI's consumer products as the polished commercial layer on top of it, similar to Google's Chrome.
That said, three assets came with Steinberger that are genuinely hard to replicate. The first is developer trust. Steinberger is not a corporate product manager who shipped an agent from inside a lab. He's an independent developer who built OpenClaw in public, bled cash to keep it running, and routed sponsorship money to dependencies instead of pocketing it. And he told Lex on camera that when it came to the acquisition talks with Meta and OpenAI, quote, I don't do this for the money. I don't give a f. That posture substantiated by the fact that he'd already had a nine-figure exit gave him authenticity with developers that no amount of marketing spend could manufacture and that it's hard to get and that it's way.
The second is architectural knowledge. Open claw is not a toy demo. It is a platform with a gateway architecture, a skills marketplace, clawhub, browser control, cron scheduling, multimodel support for claude, chat, GPT, Grock, Deepseek, and other open source LLMs, and integrations spanning a dozen messaging platforms. It runs on Mac OS, Linux, and via Docker. The security challenges it has faced and the solutions Steinberger and his community have developed represent hard one knowledge about what happens when you give an AI agent real access to real systems. That knowledge is directly transferable to whatever OpenAI is building next.
The third thing that OpenAI got ironically is community. 600 contributors, a Discord server that became a gathering point for some of the most creative and unhinged agent experiments on the internet. a global user base that includes developers building AI controlled mini breweries, smart home automations and DevOps pipelines. The Open Claw community is chaotic, inventive, and deeply invested. Exactly the kind of ecosystem Open AI needs if it wants to compete in the agent layer, and they were very wise to not claim full control of that community. It is an open- source community and would resist that kind of ownership.
The timing of all of this is not coincidental. It's not even driven by the popularity of OpenCloud necessarily, although that was a factor. Look at what OpenAI is looking at more broadly across the board in midFebruary 2026. Anthropics Claude code has hit a billion dollars in annualized revenue in just 6 months since launch. It has become the default coding tool for a generation of developers and its momentum has showed no signs of slowing despite continued strong launches from OpenAI.
OpenAI's Codeex product launched as a Mac OS app in early February was supposed to be the counter punch and positioned as a command center for agentic coding. But Codex is playing catch-up right now in a market where developer loyalty is becoming sticky and switching costs may become real. Meanwhile, Peter Steinberger has been walking around for weeks publicly describing himself as the biggest unpaid promoter for Codeex. And I have to say that makes a ton of sense.
I've been saying for a long time that deeply experienced developers like Peter get more out of codeex because codeex optimizes for correct code over long runs if you specify upfront what you need with the experience of a senior developer. That's what Peter is. He's going to get great results out of Codeex. He had been building OpenClaw using OpenAI's models all along. And he had recorded a three-hour Lex Freriedman episode, one of the most widely viewed tech podcasts on Earth, comparing GPT codeex 5.3 and Claude Opus 4.6 side by side, arguing that Codeex was reliable and efficient and that it worked well for his workflow.
His assessment was nuanced. Claude had stronger role- playinging ability and was more interactive, but was impulsive and would sometimes write code without reading context. Codeex would read a large volume of code by default before starting but was less interactive and drier in style and then it would just come back after 20 minutes of silence with the job done. The bottom line for him which is similar to what I think is that skilled developers are going to get very strong results with any top model and the differences come down to post-training goals not raw intelligence.
That kind of detailed, credible evaluation delivered on camera by a developer whose project has passed 180,000 GitHub stars heading toward 200,000 GitHub stars. It's worth more to OpenAI's developer relation strategy than any marketing campaign. Steinberger wasn't being paid to say any of this. He told Lex that joining OpenAI would feel gratifying because it would finally put a price on the work he'd done for free. He described himself as generating a tremendous amount of value for OpenAI without compensation by discussing what Codeex did for him in his role building open clock. I think he was right.
The Codeex connection runs deeper than an endorsement. Steinberger's development workflow is itself a testament to what OpenAI's coding tools are capable of doing. He's running four to 10 agents simultaneously, accumulated 6600 commits in January alone, and built most of OpenClaw's codebase by just talking to AI rather than typing. He practices what Andre Carpathy calls agentic engineer, a term he much prefers to vibe coding. His productivity on OpenClaw demonstrated at scale what codeex could enable, and that demonstration drove developers to try OpenAI's tools.
Bringing Steinberger inside means that connection becomes structural rather than accidental. The developer who proved what Codeex could do in the wild joins the company that makes codeex. Every future version of Codeex will benefit from the feedback loop of someone who has now shipped a massive 180 to 200,000 star GitHub project using it. But the developer angle is only a part of this story.
The deeper strategic logic continues to be about agents and where they're going in 2026. Open AAI has been talking about agents for months. Their responses API, their agents SDK, their agent kit, all represent building blocks for multi-step workflows. Codeex has evolved from a code completion tool into what they describe as a coding surface, combining reasoning capable models with developer tools. Sam Alman told reporters that AI models don't run out of dopamine and keep trying because they don't run out of motivation. apparently unlike developers who run on pizza.
What OpenAI has not had is a consumer-facing agent product that people actually use in daily life to manage real tasks like email, calendars, messaging, file management, all top use cases on OpenClaw today. And that's what OpenClaw delivers for OpenAI. Steinberger's stated mission at OpenAI is to build an agent that even his mom can use and is a direct acknowledgement that the current state of Agentic AI is way too technical for mainstream adoption. The gap between what OpenClaw demonstrated is possible and what a normal person can do safely on their laptop is enormous.
Closing that gap requires access to frontier models, security research, and infrastructure that a solo developer operating at a loss is not going to be able to sustain. even one with a big exit like Peter. Sam's announcement made this very explicit. As he put it, Steinberger's work on very smart agents interacting with each other to do useful things for people would become a part of our core product offerings. That's not really a compliment to Peter. That is Sam saying Peter's vision is going on OpenAI's roadmap.
You can't really understand this deal though without understanding the security crisis that has shadowed Open Claus growth from the beginning. In late January, security researcher Mav Leven of Depth first disclosed a high severity vulnerability in OpenClaw that allowed a one-click remote code execution through a crafted malicious link. The attack chain was devastatingly simple. Just clicking the link triggered a cross-sight websocket hijacking attack because OpenClaw server did not validate the websocket origin header. That meant an attacker could extract the victim's authentication token and connect to their local Open Claw gateway, disable all the safety controls and execute arbitrary commands even on instances configured to listen only on local host. If you think that sounds like they could hijack OpenClaw, that is the right reading of that bug.
Patch shipped on January 30, but the broader picture was alarming. Census identified 21,000 plus exposed openclaw instances publicly accessible on the internet, up from around a thousand just days earlier. Misconfigured instances were leaking API keys, OOTH tokens, and plain text credentials. Moltbook's database was found to expose 35,000 email addresses and 1 and a half million agent API tokens. Security firm Snick reported that 7ome% of the nearly 4,000 skills in Clawub mishandled secrets like API keys through LLM context windows. The University of Toronto issued a vulnerability notification. Hacker News published a detailed writeup.
All of this felt like a blitz of overwhelming security gaps to Peter. And Peter did his best. Version 2026.2.1 2.1 shipped on February 1st brought TLS 1.3 minimum system prompt guard rails basically a bunch of basics for security and Peter kept going the next version shipped February 7th with an added code safety scanner and added support for new models and then version 2026.2.12 2.12 released February 12th, just 2 days before Steinberger announced that he was going to open AI. That was the big one. 40 plus dedicated security patches addressing prompt injection, rce and browser control, unauthenticated configuration tampering, and a bundled hook identified as quote soul evil that had inadvertently remained in the codebase.
The timing for all of this security work is significant. Steinberger shipped the most comprehensive security overhaul in Open Claus history in the same week that he was finalizing his decision to join OpenAI. He did not leave the project in a vulnerable state. He fortified it and handed it to a foundation. I think this also explains a part of OpenAI's calculus. The security challenges OpenClaw faced are not unique. They are inherent to the category. Any company shipping autonomous agents that can access email, execute shell commands and manage calendars is going to face these problems. Steinberger has now lived through them and Steinberger has developed practical responses to so many of them even operating as a solo developer powered by agents. That kind of scars on his hands experience is operationally valuable to OpenAI in a way that no amount of theoretical security research is going to be able to replicate.
So here's the question for all of us. What changes for Open Claw now? For the 600 contributors and hundreds of thousands of users, the immediate answer is both not much and everything. Open Claw will move to a foundation structure. It will remain open source. It will continue supporting multiple models, not just Open AIs. And Steinberger has been explicit that the project should grow to support even more model providers and companies. OpenAI has committed to continuing its sponsorship of the project regardless.
The Chrome Chromium analogy that Peter used talking to Lex in his interview is instructive here, although perhaps not in the way he intended. Chrome is built on the open-source Chromium project, but Google's influence on Chromium's direction is dominant. Google engineers contribute the majority of commits, set architectural priorities, and the features that make it into Chrome shape what Chromium becomes. Independent Chromiumbas based browsers like Brave or Edge operate within a framework largely defined by Google's priorities.
Obviously, the risk for OpenClaw is similar. With Steinberger now inside Open AI, the project's founder and most prolific contributor will inevitably be influenced by his employer's priorities. Features that align with OpenAI's product roadmap may get faster attention. Features that compete with OpenAI's offerings. features. The foundation structure is designed to mitigate this. But foundations are only as independent as their governance allows. In the details of Open Claus Foundation, like its board composition, its funding sources, its contribution policies, those have all not been announced.
There's also a practical question about the 3,000 plus open poll requests that Steinberger mentioned before the deal. He committed to processing them regardless of his decision. But a solo developer becoming a full-time OpenAI employee will necessarily have less discretionary time for open-source maintenance. The OpenCloud community is going to need to develop its own leadership bench.
The upside for OpenClaw users is real, too. Open AAI has resources, compute, security teams, model access, infrastructure that an independent project cannot hope to match. If OpenAI follows through on its commitment to sponsor the project in a meaningful way, which is an if, Open Claw could get more robust faster than it would have as a one-person operation hemorrhaging cash. The security hardening alone could benefit enormously from access to OpenAI security research.
So, here's the multi-billion dollar question. Now, where does OpenAI go? Sam's phrasing that Steinberger's work would quickly become core to OpenAI's product offering suggests a very specific product direction to me. OpenAI appears to be building very aggressively toward a consumer agent product that goes well beyond Chad GPT's current capabilities.
Consider what OpenAI now has in the portfolio. Codex handles coding agents. Chad GPT handles conversational AI. The responses API, agents SDK, and agent kit all provide developer infrastructure for multi-step workflows. What's missing is a persistent always on personal agent that manages the messy cross-platform reality of how people actually use their computers and phones. The email triage, the calendar conflicts, the Slack follow-ups, the file organization, the proactive task management. That is what OpenClaude demonstrated was possible. And that's what Steinberger says he wants to build at Open AI, an agent his mother can use.
The technical challenges are formidable here. Open Claw Security Crisis proved that giving an AI agent broad access to a user's digital life creates an attack surface that current security models really struggle to contain. Steinberger's own maintainer, known as Shadow, warned on Discord that if someone can't understand how to run a command line, this project is far too dangerous to use safely. Making that same capability safe for non-technical users requires solving problems like sandboxing, permission management, data sovereignty, and model reliability requires solving problems like sandboxing, permission management, data sovereignty, and model reliability that are at the frontier of what anyone in the industry knows how to do.
Sam's mention of smart agents interacting with each other also signals an interest in multi- aent architectures, systems where specialized agents collaborate on complex tasks. This aligns with what OpenAI demonstrated in its harness engineering case study where a team of three engineers used codeex to produce over 1500 pull requests across a million line codebase with zero human written code. The extension of that model from coding to personal productivity seems like a very natural next step and the competitive implications here are going to be significant.
Anthropics cloud code dominates the developer tools market right now. Google is investing in Gemini based agent capabilities for the consumer. metacorted Steinberger personally with suck reaching out via WhatsApp and reportedly spending 10 minutes arguing with Peter over whether claude or chat GPT was the better coding model. Microsoft which invested heavily in OpenAI has its own agent ambitions through co-pilot. Apple has been quiet but has been benefiting from that run on Mac minis that Peter sparked through the open claw project. Steinberger's hire gives OpenAI a credible claim to win in the personal agent space. Not because of any proprietary tech, but because of the proven execution Peter brings. He built something 200,000 developers want to use. He did it in just a few months. And he did it in a way that generated the kind of organic enthusiasm that no marketing budget can buy.
Let's step back and look at the bigger picture here. Beyond OpenAI, beyond OpenClaw, Peter told Lex that OpenClaw style agents would kill 80% of apps. His logic is pretty simple. Every app is just a slow API to what the user wants. And an agent that already knows your location, your sleep patterns, your stress levels, and your calendar doesn't need you to open a separate app to handle things like fitness tracking, food ordering, and scheduling. It will just do it.
That prediction may prove aggressive on the timeline and conservative on the scope. What I mean by that is that we don't know when that's going to actually catch. I know OpenAI is going to be eager to have personal agents everywhere, but product market fit is notoriously difficult to get with consumers. Open AAI caught Chat GPT lightning in a bottle once. It's not yet clear if they can do it twice with a personal agent.
Regardless, the more fundamental shift is not about replacing apps, but about changing the interface layer between humans and software. For 30 years, the dominant paradigm has been graphical user interfaces, icons, menus, buttons, and for the past 15, it's been touch interfaces on mobile. What OpenClaw demonstrated imperfectly and a great personal cost to its creator is a third paradigm, delegation. You don't tap an icon, you don't type a query, you just tell the agent what you want done, and it figures out which APIs to call, which tools to use, which steps to take.
The fact that this paradigm emerged not from a corporate lab but from a single developer's living room in Vienna is itself instructed. It suggests that the hard problem in Agentic AI is not primarily one of model capability. The underlying LLMs were already good enough. The primary challenge is one of integration, persistence, and the willingness to give an AI system real access to real things. Steinberger's contribution was not a new algorithm. It was glue code, architectural decisions, a messaging interface, and the audacity to let an agent modify its own source code.
Now, that legacy of Audacity lives on inside OpenAI. And the question is whether it survives the transition from an indie hacker project into a corporate product and whether the foundation model preserves enough independence to keep the open- source open claw community invested. The lobster has molted for the last time. What it grows into next depends on whether the new shell fits.