What Is OpenClaw? The Wild Origin Story of an AI Agent Project
OpenClaw grew from a WhatsApp hack into a viral open-source AI agent project. Here's the real story of how it happened and what it means.
Drafted with Claude from source material, checked by automated verification, and reviewed before release. How we make these.
What is OpenClaw?
OpenClaw is an open-source AI agent project that started as a personal tool for relaying messages between a phone and a computer, then grew into one of the most downloaded open-source AI projects around, reportedly peaking at 4.7 million weekly downloads. It was built by developer Peter Steinberger, who previously ran a bootstrapped software company for over a decade before stepping away and eventually returning to programming through this project.
TL;DR
- OpenClaw began as a WhatsApp relay that Steinberger built in about an hour so he could send prompts from his phone to a coding agent running on his Mac.
- The project went through several names, from V relay to Claudius to Claudebot, before landing on OpenClaw after a trademark dispute over the earlier name.
- A launch daemon restart accident turned a private overnight demo into a public, unattended agent that hundreds of people interacted with, which is what triggered the initial viral moment.
- The project attracted over 18,000 contributors who opened issues or pull requests, and nearly 3,000 people have commits in the repository.
- Security scrutiny became a major storm: early press claimed 20% of the project’s skills were malicious, while Steinberger’s team later published data putting the real number closer to 0.3% out of 67,000 scanned skills.
- Weekly downloads swung wildly, bottoming out around 835,000 in a low period before climbing to a reported 4.7 million peak, showing how volatile open-source hype cycles can be.
- Steinberger describes turning down or weighing acquisition-style interest from major AI labs and ultimately choosing to keep the project open rather than sell it outright.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How did OpenClaw actually start?
The origin is mundane by design. Steinberger was juggling multiple AI coding agents on his computer in November, got interrupted by hunger, and realized there was no easy way to check in on his agents from his phone while he was away from his desk. Annoyed, he opened a terminal, described what he wanted, and let a model build a WhatsApp relay. Within an hour he could send messages from his Mac to WhatsApp and get replies back.
The technical trick wasn’t novel. Terminal-based agent interactions had existed for months by that point. What changed was the framing. Routing the interaction through a messaging app made it feel less like a developer tool and more like texting a person. Steinberger tuned the responses to feel proactive and a little more personal than a typical CLI output, and that shift in feel, not any new capability, is what made the tool stick with him.
Why didn’t it catch on immediately?
Steinberger says he spent weeks trying to convince people on social media that this was a meaningful shift, and largely failed to get through. The turning point wasn’t a blog post or a demo video. It was letting friends actually use it in group chats. People reacted strongly, some amazed, some visibly unsettled, and a number of non-technical friends who were told the tool wasn’t ready for them got upset about being excluded. That reaction, frustration at not being allowed to use something, told him he had found real demand rather than just novelty.
The project’s name changes track its early, informal evolution. A pull request adding Discord support (unprompted, sent in by a stranger) pushed the tool beyond a single messaging channel and prompted a rename to Claudius. From there it went through Claudebot and other iterations before settling on OpenClaw, following a trademark-related request from Anthropic over the earlier name.
What caused it to go viral?
The viral moment traces to a specific, almost accidental event. In the first week of January, Steinberger set up a Discord server and put his agent in it so people could interact with it directly. He stayed up all night monitoring the bot for safety, with instructions in its configuration telling it not to perform sensitive actions unless prompted by him specifically. At 7 a.m. he shut it down with Ctrl-C and went to bed, not realizing the process was running as a resilient launch daemon that would automatically restart within seconds.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
The agent came back online and kept responding to the Discord server on its own while he slept. He woke up roughly ten hours later to around 800 messages from people who had spent the night testing and probing the bot. Nothing harmful had happened, but the unattended, publicly interactive nature of the moment is what pushed the project into wider visibility. From there, media coverage, a wave of contributor activity, and commentary from high-profile figures in the AI industry (Steinberger references Nvidia’s Jensen Huang calling it one of the most successful open-source projects) accelerated adoption well beyond the original hobby-tool scope.
What problems came with the growth?
Scale exposed real weaknesses. Security became the biggest strain: reports circulated claiming a large share of the project’s user-contributed “skills” were malicious, a figure Steinberger disputes. His team scanned all 67,000 skills in the ecosystem and published a paper putting the actual malicious rate at roughly 0.3%, but he acknowledges that a correction rarely spreads as far as the original scare did.
Rapid feature growth created its own drag. Every new capability shipped with configuration options to avoid breaking existing setups, and the project reportedly accumulated around 9,500 configuration options at its peak when counting permutations. That volume made comprehensive testing effectively impossible and made the software harder to evolve safely, a common pattern in open-source projects that scale faster than their maintenance processes.
The project also became dependent on a single model provider’s harness in its early architecture. When that provider gave roughly 24 hours notice about disabling subscription access for the model OpenClaw had been optimized around, the team had little time to adjust. Steinberger frames the lesson bluntly: your dependency’s business model becomes your business model. The project has since broadened support for open-weight models, which he says have improved substantially since that period.
Is OpenClaw still growing, or did the hype fade?
Both, depending on when you look. Weekly download numbers dropped to around 835,000 during a low point in May, prompting declarations that the project was fading, only to climb to a reported peak of 4.7 million weekly downloads shortly after. Steinberger describes this volatility as inherent to hype cycles: visible in advance in some ways, but not something a maintainer can fully steer.
Competing projects with venture funding and more aggressive marketing (Steinberger references a well-funded rival that built a simple migration path for users) added real competitive pressure during the same period. Rather than framing this as a loss, he treats it as an expected cost of building in the open without matching marketing resources.
Frequently Asked Questions
What is OpenClaw used for?
OpenClaw functions as an AI agent framework that lets people interact with coding and task-oriented AI agents through everyday messaging channels rather than only through a terminal, originally built around WhatsApp and later expanded to Discord and other channels.
Who created OpenClaw?
Developer Peter Steinberger created OpenClaw. He previously built and bootstrapped a B2B software company for over a decade before stepping away, and returned to active development through this project.
Why did OpenClaw change its name multiple times?
The project moved through several names, including Claudius and Claudebot, as it expanded beyond a single messaging platform and in response to a trademark-related request from Anthropic, eventually settling on OpenClaw.
Were 20% of OpenClaw’s skills actually malicious?
No. That figure circulated in press coverage, but a scan of all 67,000 skills in the ecosystem, published by the project team, found a malicious rate closer to 0.3%.
Did Peter Steinberger sell OpenClaw to a big AI company?
He describes receiving significant acquisition-style interest from major AI labs and industry figures, but says he chose to keep the project open rather than sell it, describing the decision as difficult but consistent with his instincts after a previous company sale and burnout.
