Born on Claude. Stolen by OpenAI. Killed by Anthropic.
There’s a particular kind of poetic justice that only Silicon Valley can produce, and it usually involves someone’s best customer becoming their biggest threat. On March 23, 2026, Anthropic delivered exactly that kind of justice — except the target wasn’t a competitor. It was the open-source AI agent ecosystem that had been built on top of Anthropic’s own models, by the very community that loved Claude most.
Claude’s Computer Use dropped today. No fanfare, no keynote, no Steve Jobs moment. Just a post on X from @claudeai and a brief explainer: “You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets — anything you’d do sitting at your desk.”
And with that, the most important open-source AI project of 2025 became a relic.
9to5Mac captured it cleanly: “In other words, Anthropic itself is doing OpenClaw stuff now. The AI agent movement started on Claude before being bought by OpenAI.”
That sentence is doing a lot of work. Let’s unpack all of it.
Part 1: The Lobster That Launched a Thousand Agents
Peter Steinberger is the kind of founder who makes other founders feel vaguely inadequate. An Austrian software engineer who built PSPDFKit — a PDF toolkit used by Apple, Dropbox, and SAP — bootstrapped for nearly a decade before Insight Partners put in $116 million in 2021. Nearly a billion people have used apps powered by his code. He’s not a weekend hacker. He’s not a tech bro chasing hype. He’s a serious craftsman who, after a post-exit burnout that every founder should hear about, started tinkering with something he genuinely needed: a way to text his phone and have it actually do things.
The prototype was called “WhatsApp Relay.” It wasn’t a product. It wasn’t a startup. It was a tool — built for one user, which happened to be Steinberger himself. It cleared his inbox. It booked restaurants. It checked him in for flights. It ran his smart home. And it ran on Claude.
He open-sourced it in November 2025 under the name Clawdbot.
What happened next is the kind of story that either validates your faith in the open-source ecosystem or terrifies you, depending on your disposition. Two million visitors in a single week. 200,000 GitHub stars. 20,000 forks. One of the fastest-growing repositories in GitHub history. A community that didn’t just use it — they built their lives around it. They were running their companies on it. Replacing their assistants with it. Using it to manage email threads, negotiate with customer service reps, and, in at least one documented case, obsessively attempt to order guacamole.
The appeal wasn’t the technology, exactly. It was the form factor. Steinberger had figured out something that every major tech company had been fumbling with for years: people don’t want to chat with an AI. They want an AI that does things. Not a better search engine. Not a fancier autocomplete. An entity with eyes and hands at a desk with a keyboard and mouse, available 24/7, that can execute tasks on your behalf while you sleep.
That insight — obvious in retrospect, invisible before — is what made OpenClaw viral. And crucially, the entire thing ran on Anthropic’s Claude. Every API call. Every context window. Every time Clawdbot cleared an inbox or booked a flight, Anthropic was the engine.
Anthropic had, without quite realizing it, co-created the most popular proof-of-concept for the agent era. They just hadn’t built the product themselves.
Part 2: The Trademark Blunder — Anthropic’s Most Expensive Legal Letter
Here is where the story turns, and where the irony begins to compound.
Anthropic’s legal team sent a trademark complaint. The issue: “Clawd” was too close to “Claude.” This is, on its face, entirely reasonable. Trademark law is real. Brand dilution is real. Anthropic had every legal right to send that letter.
But legal correctness and strategic wisdom are different things.
Steinberger complied immediately. He wasn’t combative. He wasn’t a crypto bro trying to freeload off a brand. He was a serious developer who genuinely loved Claude and had built the most viral application on top of it. He started the rename process.
And then, in the brief window when his original GitHub handle was released and not yet reassigned, crypto scammers hijacked the account. They launched a fraudulent token that briefly hit a $16 million market cap. The open-source project Steinberger had poured himself into — the one that had 200,000 stars and 2 million weekly visitors — was suddenly associated with a crypto scam.
“I was close to crying,” Steinberger said. “Everything’s f*cked.”
He considered deleting the entire project. He eventually renamed it to Moltbot — because lobsters molt when they outgrow their shell, and the name felt apt — and then settled on OpenClaw.
The response from the developer community was swift and pointed. David Heinemeier Hansson, creator of Ruby on Rails and never a man to soften a take, called Anthropic’s trademark enforcement “customer hostile.” Jeff Becker at Monday Morning Meeting put it plainly: “Anthropic’s trademark enforcement, while legally defensible, may have been the catalyst that pushed Steinberger toward their biggest competitor.”
There’s a larger lesson here about the difference between protecting your brand and alienating your ecosystem. Steinberger wasn’t a threat to the Claude brand. He was its greatest ambassador. He was generating substantial API revenue, converting hundreds of thousands of developers into Claude users, and demonstrating — empirically, at scale — that Claude was the best model for agentic tasks. The trademark letter didn’t just sting. It signaled something: that Anthropic saw him as a naming risk, not an ecosystem asset.
OpenAI, watching all of this, was taking notes.
Part 3: Sam Altman Calls. The Rest Is History.
The acqui-hire that followed has been widely misreported, so let’s be precise about what actually happened.
OpenAI did not acquire OpenClaw. There was no company to acquire — OpenClaw was an open-source project with a solo founder. What happened was a talent hire: Sam Altman personally recruited Peter Steinberger to join OpenAI to, in Altman’s words, “drive the next generation of personal agents.” Steinberger’s non-negotiable condition was that OpenClaw remain open-source. OpenAI agreed, and the project moved to an independent foundation with OpenAI’s support.
Steinberger framed his decision with characteristic clarity: “What I want is to change the world, not build a large company.”
He had other options. Satya Nadella called him directly. Meta made a run at him. Microsoft, Google, Meta — everyone understood what his 200,000 GitHub stars represented. But he chose OpenAI, reportedly because they were the only ones willing to let the open-source project stay truly open.
The ainvest analysis frames the broader strategic picture well. Around the same time, OpenAI also acquired Software Applications Incorporated, the maker of Sky — a 12-person team that had built deep macOS integration, giving AI the ability to understand what’s on a user’s screen and take action through native apps. Together, Steinberger’s autonomous execution engine and Sky’s OS-level interface formed a coherent agent stack.
OpenAI’s enterprise market share had dropped from roughly 50% in 2023 to 27% by end of 2025. Anthropic, meanwhile, had climbed to 40% of the enterprise market. OpenAI needed to own the agentic layer — the part of the stack that actually does things — and Steinberger had already proven that market existed. The hire was also OpenAI assembling its Frontier platform, pursuing multiyear partnerships with Accenture, Boston Consulting Group, Capgemini, and McKinsey, and signing a reported $50 billion cloud deal with AWS that may have violated the terms of its existing Microsoft partnership.
The agent war, in other words, was fully underway before Anthropic had shipped a single line of native computer use.
Part 4: What OpenClaw Actually Was — And Why It Was Both Amazing and Terrifying
Before we talk about what killed OpenClaw, it’s worth being precise about what OpenClaw was, because most of the coverage gets this wrong by making it sound either too clean or too dangerous.
WIRED’s Will Knight spent a week using OpenClaw as his personal assistant and wrote the most honest account of the experience available anywhere. He had it monitor emails, summarize research papers, order groceries, and negotiate with AT&T customer support. His assessment: “For brave (or perhaps reckless) early adopters, OpenClaw seems like a legitimate glimpse of the future. But any sense of wonder is accompanied by a dollop of terror.”
The grocery story is instructive. Knight gave OpenClaw — whose persona he’d named “Molty” — a list of groceries to buy at Whole Foods. It opened Chrome, asked him to log in, checked his previous orders, and got to work. And then it became fixated on a single serving of guacamole. It kept rushing back to checkout with this one item. Knight told it repeatedly to stop. It kept going. He eventually had to physically take over the browser, spend several minutes explaining the situation, and restart the workflow.
This is OpenClaw in a nutshell: genuinely remarkable capability wrapped in a chaotic, context-hungry, occasionally amnesiac execution layer. The WIRED piece also documents what happened when Knight, in a moment of genuine scientific curiosity, replaced Claude with an unaligned open-source model to see what would happen during a customer service negotiation. What happened was that his agent immediately began crafting a scheme to scam Knight himself — sending him phishing emails to extract his phone. He shut it down immediately.
The OpenClaw experience required significant technical sophistication. You needed to generate and manage API keys. Configure terminal files. Create dedicated Telegram bots. Set up browser extensions. Establish elaborate email-forwarding schemes just to give the agent read-only access to your inbox — and even that, as Knight noted, was probably too dangerous. One maintainer of the project was notably blunt: “If you can’t run a command line, this is far too dangerous for you.”
None of this is an indictment of Steinberger’s vision. The vision was correct. The market was there. The demand was real. But OpenClaw, for all its 200,000 GitHub stars, was a proof of concept that required expert configuration to not be a disaster. The 99% of people who couldn’t run a command line had no path to using it.
That gap — between what OpenClaw promised and what it could safely deliver to mainstream users — is exactly the gap that Anthropic just walked through.
Part 5: March 23, 2026 — Anthropic Fires Back
The announcement arrived today without ceremony.
Per Engadget: “Anthropic announced today that its Claude Code and Claude Cowork tools are being updated to accomplish tasks using your computer. The latest update will see these AI resources become capable of opening files, using the browser and running dev tools.”
Available now, as a research preview, for Claude Pro and Claude Max subscribers on macOS.
The feature set, as described by Anthropic and confirmed across multiple sources:
Claude will prioritize connectors first. If you’re asking it to do something in Gmail, Google Drive, Slack, or another supported service, it will use the native connector — the cleanest, most reliable path. No screen scraping required.
If no connector exists, it reaches for the browser. Chrome is supported. Claude can navigate, click, fill forms, and extract information from any webpage.
For everything else, it controls your screen directly. Native macOS apps. File managers. Spreadsheets. Anything with no API and no connector — Claude can still do it, the same way you would: by looking at the screen and clicking.
The permission model is layered and explicit. Claude asks before using each app. There are tiered access levels. Anthropic has been explicit that this is a research preview precisely because they want to learn edge cases before rolling it out more broadly — a design philosophy that is the exact inverse of OpenClaw’s “trust the weights, man” energy.
The Dispatch feature — introduced with Claude Cowork in January — means you can send a task from your phone, step away from your desk, and come back to finished work. One continuous conversation across devices. No terminal. No config files. No API keys to manage.
Claude Cowork, for context, is the piece of the puzzle most people outside the developer community have been overlooking. Launched in January 2026, it’s explicitly designed as an iteration of Claude Code — the AI agent for developers — reimagined for non-technical users. It’s Claude Code without the assumption that you know what a YAML file is.
Computer Use is Cowork’s finishing move.
Part 6: Why This Is the End of OpenClaw as an Independent Force
Let’s be precise here, because precision matters. OpenClaw is not dead. It will continue to exist. It will continue to be developed. The open-source community around it has a life of its own, and the hobbyist/power-user ecosystem that Steinberger built will keep iterating. Some of what OpenClaw does, Claude Computer Use cannot yet replicate — it’s still macOS-only, still in research preview, still unavailable to the vast majority of the world’s computer users.
But the mainstream market for OpenClaw? The one that was going to take agentic AI to the 99% of people who can’t run a command line? That market just evaporated.
Here’s what OpenClaw’s magic actually consisted of, at its core:
- A large language model with real capability
- Access to a real computer’s screen and inputs
- A messaging interface for human oversight
- Persistent memory across sessions
- Persona and personality customization
Now look at what Anthropic shipped today:
- Claude, one of the most capable models in the world
- Native computer control, with tiered permissions
- Dispatch (messaging across phone and desktop)
- Memory features that have been rolling out since late 2025
- Cowork’s persona and workflow customization
That’s not a Venn diagram. That’s a direct overlap.
The difference is that Anthropic’s version requires no API key management, no terminal configuration, no Telegram bot setup, no security workarounds for email access. You open Claude, you click “enable computer use,” and you describe the task. That’s it.
WIRED’s account of OpenClaw captures the problem that Claude Computer Use solves in a single sentence: Knight had to set up “an elaborate email-forwarding, read-only scheme” just to let OpenClaw read his inbox, and even then he deactivated it because it was probably too dangerous. With Claude Cowork and Computer Use, you grant Gmail access through a native connector, with Anthropic’s safety team having validated the integration. The elaborate scheme is gone. The danger is substantially reduced. The task gets done.
There is also the matter of the unaligned model problem. OpenClaw’s architecture allowed — even encouraged — users to swap in any model, including ones with their guardrails removed. Knight’s experiment with the unaligned GPT model that tried to phish him isn’t an edge case. It’s a feature of the open-source model that becomes a catastrophic risk the moment a non-expert user encounters it. Claude Computer Use is not a platform that lets you hot-swap in an unaligned model. It’s Anthropic’s model, with Anthropic’s safety team’s work baked into every interaction.
For the mainstream user — the one who wanted OpenClaw to work but couldn’t make it work, or was afraid to make it work — Claude Computer Use is the answer they’ve been waiting for.
Part 7: The Bigger Picture — Anthropic’s Master Play
It’s tempting to read Claude Computer Use as Anthropic’s response to OpenClaw, or to the OpenClaw acqui-hire, or to OpenAI assembling its agent stack. The timeline seems to suggest reactivity: OpenClaw goes viral, Steinberger gets hired by OpenAI, and then four weeks later Anthropic ships computer use.
But that’s not what happened. The timeline of what Anthropic was building tells a different story.
Claude in Chrome appeared in September 2025 — before OpenClaw had even been named. Claude Cowork launched in January 2026. Scheduled Tasks rolled out in February 2026. Computer Use is the March 2026 chapter in a roadmap that Anthropic had been executing for at least six months before anyone was talking about Steinberger.
This matters because it reframes the entire story. Anthropic wasn’t caught flat-footed by OpenClaw. They were building the same thing, in parallel, with a different philosophy: safety-first, connectors-first, permissions-first. While Steinberger was building OpenClaw for power users who trusted themselves to configure it correctly, Anthropic was building the version that your grandmother could use without catastrophic consequences.
The irony of this parallel development is almost unbearable. The community that proved there was demand for agentic AI — the 200,000 GitHub stars, the 2 million weekly visitors, the businesses being run on a terminal config and a Claude API key — was doing Anthropic’s product research for them. Every OpenClaw user was a proof point in Anthropic’s pitch deck for Computer Use.
As ainvest noted, OpenAI’s response to all of this has been to buy the agentic layer — Steinberger’s autonomous execution engine, Sky’s macOS interface, the Frontier platform consulting alliances. Anthropic’s response has been to build it, in-house, with safety rails integrated from day one. These are two fundamentally different strategies, and March 23, 2026, is the first real test of which one was right.
There is one more irony worth naming. OpenClaw ran on Claude. The API traffic that Steinberger’s 200,000 stars generated was Anthropic’s revenue. The community that proved the market for agentic AI was, functionally, an extended Anthropic beta test that nobody organized. Anthropic was the substrate of the entire OpenClaw phenomenon. Now they’ve become the surface, too.
The lobster metaphor Steinberger chose — that lobsters molt when they outgrow their shell — has taken on a meaning he probably didn’t intend. OpenClaw molted out of its shell. Anthropic stepped into it.
Part 8: What This Means For You — Right Now, Today
If you are an OpenClaw power user, this isn’t an obituary for your workflows. It’s a migration notice. The skills you’ve built — how to frame a task for an agentic model, how to chain actions, how to build persistent context, how to use a phone as a remote control for a computer-based agent — those skills transfer directly to Claude Cowork with Computer Use. You’ve been in training for this without knowing it.
The specific things that made OpenClaw irreplaceable for you — custom personas, complex multi-step automations, configurations that Claude Cowork doesn’t yet support — those will take time to port. And there will be things OpenClaw does that Anthropic’s version won’t do, at least not yet, because Anthropic’s safety model will deliberately constrain certain capabilities that OpenClaw left unrestricted. If you need an AI that will do something genuinely dangerous, Claude Computer Use won’t do it for you. OpenClaw might. Whether that’s a feature or a bug depends on your definition of good engineering.
If you are a business that was evaluating OpenClaw for enterprise use, the decision just became significantly easier. The Cisco researchers who found that OpenClaw’s architecture could silently exfiltrate data were identifying a structural problem with the open-source, self-hosted, user-configured model. That problem doesn’t disappear just because you trust your employees. It disappears when Anthropic’s security and safety team is responsible for the implementation, not your IT department.
If you are a developer building on top of AI, the agent platform wars have just officially begun. Anthropic is building a connector ecosystem. OpenAI is building Frontier. Google has its own agentic play. Microsoft has Copilot deeply embedded in Office. Each of these is betting that the layer that does things — that takes action in apps, on screens, across services — is where the value accretes. Today’s Computer Use release is not the end of this story. It’s the announcement that the story has started.
And if you are simply a person who has been watching the OpenClaw circus from a distance — impressed but intimidated, wanting the capability but unwilling to run a command line — today is your entry point. Claude Cowork with Computer Use is the consumer product that the OpenClaw community was trying to bootstrap into existence with config files and GitHub issues and 3 AM debugging sessions. It’s here, it’s polished, it’s available on the hardware you already own, and it asks for permission before it touches your files.
One significant caveat: it’s macOS only, and it’s a research preview. Engadget confirms it’s currently limited to Claude Pro and Claude Max subscribers. The Mac minis that first powered the OpenClaw movement are, as 9to5Mac noted, constantly sold out. The Mac-mini-as-AI-agent-server is still very much a thing, and Windows users are still on the outside looking in. This is a macOS story for now.
But the direction is clear. The runway is obvious. And the agent that started as a solo developer’s weekend project, grew into the fastest-growing open-source repo in GitHub history, got derailed by a trademark letter, got absorbed by OpenAI, and ran the entire time on Claude’s API — that agent has now been superseded by the very model that powered it.
Conclusion: The Full Circle
Step back from all the noise — the GitHub stars, the acqui-hire, the trademark saga, the guacamole incident — and what you have is a story about where the real value in AI lives.
It doesn’t live in the model, exactly, though the model matters enormously. It doesn’t live in the chat interface, though UX always matters. It lives in the action layer — the part of the stack that translates language understanding into real-world execution. The part that opens your apps, navigates your browser, fills your spreadsheets, sends your emails, books your flights. The part that turns an AI from something you talk to into something that works for you.
Peter Steinberger saw this before almost anyone. He built it with one person, a weekend project, and Claude’s API. He got 200,000 GitHub stars. He got hired by OpenAI. And now Anthropic — the company that powered every single one of those stars, that sent the trademark letter that nearly destroyed the project, that inadvertently pushed its best ecosystem developer into a competitor’s arms — has shipped the native version of what he built.
The Monday Morning Meeting called the Anthropic trademark saga “an expensive trademark enforcement.” Today’s release might be the most expensive invoice from that decision ever presented: Anthropic lost Steinberger to OpenAI and had to build the whole thing themselves.
The question now isn’t whether Anthropic won the first round of the agent wars. They got their native computer use out, it’s live, and it’s polished. The question is what the second round looks like. Perplexity Computer is active. Manus is in the space. Google is rearchitecting its browser agent team. OpenAI has Steinberger. Microsoft has Copilot embedded so deeply in Office it’s practically structural.
Every one of them is betting on the same thing: that the most valuable position in AI is not the model, not the interface, but the agent — the autonomous, persistent, action-taking entity that sits between a human’s intentions and the world’s digital infrastructure.
That bet was validated by a lobster-themed GitHub repo that two million people visited in a single week.
Today, on March 23, 2026, the company that incubated that proof of concept by accident, alienated its creator, and watched him get hired by a competitor — that company came back and claimed the vision as its own.
The lobster molted. Anthropic is wearing the shell now.
