How OpenClaw Became the Hottest AI Agent in the World

追加: 2026-05-17

[00:00:00]OpenClaw went from basically unknown to more than two million visitors in a week, crossed 100,000 GitHub stars at crazy speed, and got so big so fast that even Anthropic said tools like it were putting an outsized strain on their systems.

[00:00:16]That is the kind of momentum you expect from a viral app, not from a technical open-source AI agent. And yet, OpenClaw pulled it off. That is why OpenClaw matters because this is not just another chatbot. It is part of the next wave in AI software that does things for you, not just talks to you. It can read your messages, connect to your apps, click through workflows, and in some cases run commands on your machine. That makes it exciting, useful, and honestly a little terrifying. And it also raises a business question. If the product is open source, if the founder gets hired by open AI, if Meta buys a neighboring company instead of the product itself, then where is the money? That is the story here. Not just what OpenClaw is, but why it became important so fast, why the industry started chasing it, and why its biggest lesson may actually be about failure.

[00:01:10]The founder behind OpenClaw is Peter Steinberger, and his story does not look like the usual dropout founder raises venture money script.

[00:01:20]He grew up in rural Austria, got obsessed with computers as a teenager, studied software engineering at the Vienna University of Technology, worked as a senior iOS engineer in Silicon Valley, and even taught mobile development back at his university. Long before Openclaw, he had already built a serious software career.

[00:01:41]For about 13 years, Steinberger built a successful PDF software business. One Fortune syndicated profile put the contrast perfectly. He spent 13 years building a company around PDFs and then built the prototype that would eventually overshadow that work in about an hour. That line matters because it tells you OpenClaw did not come from a lab with a giant research team. It came from a seasoned software builder who had already spent years learning how real products break in the real world.

[00:02:10]OpenClaw itself appears to have started in late 2025 as a personal side project.

[00:02:16]The bot first gained popularity after appearing in November, while OpenClaw's own blog formally introduced the platform publicly in January 2026.

[00:02:26]The basic idea was deceptively simple.

[00:02:30]Take a strong language model, connect it to the tools people already use, and let it act as a kind of personal digital employee.

[00:02:39]Steinberger built the first version to work through chat apps and to run on the user's own machine rather than in a closed SAS environment.

[00:02:47]That design choice is a big part of why OpenClaw took off. It felt personal, hackable, and much closer to the original spirit of open-source software than the polished assistance coming out of the big labs. So, what exactly is OpenClaw? The official description from the project says it is an open agent platform that runs on your machine and works through the chat apps you already use. WhatsApp, Telegram, Discord, Slack, Teams, wherever you are. In simple terms, it turns a language model into an operator. Instead of just answering a question, it can be connected to files, apps, terminals, calendars, browsers, and AP is so it can actually carry out tasks. That is the jump from chatbot to agent. A chatbot gives you information.

[00:03:33]An agent can take action. So if you ask a chatbot, "What flight should I take?"

[00:03:39]It gives you suggestions. If you ask an agent, "Check me in for tomorrow's flight, move my morning meeting," and send a message to my team. The point is that it should go do those things.

[00:03:50]Reporting around OpenClaw's rise highlighted use cases like managing email, calendars, insurance tasks, and flight check-ins. That is why the product caught fire. It made AI feel less like a toy and more like labor.

[00:04:05]The real value is not that OpenClaw is smarter than every other model. The real value is that it wraps intelligence inside a usable system. It is a tool that can connect hardware and software tools and learn from the resulting data with much less human intervention than a typical chatbot.

[00:04:22]NVIDIA's Jensen Huang went even further and compared the phenomenon to an operating system for a new age of AI agents.

[00:04:30]That framing helps explain the excitement. OpenClaw was not just another model rapper. People started seeing it as a way to orchestrate digital work and that is also what separated it from competitors.

[00:04:43]Most mainstream assistants still lived inside walled gardens.

[00:04:47]Openclaw lived on your machine, used your infrastructure, and could be wired into your tools. That made it more flexible and for technical users, much more powerful. Now, here is where OpenClaw becomes tricky to analyze as a company.

[00:05:03]OpenClaw did not scale like a normal ventureback back SAS company with a clean pricing page, a disclosed ARR number, and a polished B2B sales team. Public reporting does not show a traditional venture fundraising story for OpenClaw itself. In fact, the clearest business milestone was not a financing round, but OpenAI hiring Steinberger and helping move OpenClaw into a foundation structure so the project could stay open- source. In other words, the product's influence exploded before its corporate structure really did. That means the money around OpenClaw showed up in a different place, the ecosystem.

[00:05:41]By late February 2026, trackers cited by several crypto and market outlets said that roughly 129 startups built around OpenClaw had generated about $283,000 in revenue over 30 days with the top project doing around $50,000 monthly.

[00:05:59]Those businesses were selling hosted versions, setup services, integrations, tools, wrappers, and other convenience layers around the open-source core.

[00:06:09]That is a small number compared with enterprise software giants, but it is a real number and more importantly it shows where monetization likely lives.

[00:06:19]Not in owning the protocol, but in selling the easier way to use it. So if you want the simplest answer to how does OpenClaw make money, it is this. By itself, OpenClaw has looked more like an open-source platform than a mature operating company. While the money has flowed into the surrounding layer of hosting, services, support, and enterprise tooling.

[00:06:43]That is why it spread so fast. People did not have to wait for a giant company to productize every use case. They built their own. But that loose model created a second problem. It made it much harder to control quality, security, and reliability. And that is where the failure story begins.

[00:07:02]OpenClaw's promise was that AI agents could finally do real work. Its biggest weakness was that real work means touching real systems. And once an AI agent can touch real systems, mistakes get expensive fast. Let's start with reliability. In demos, agents look magical. They summarize email, move meetings, book travel, compare prices, and open websites in sequence. But the more steps you give them, the more room there is for drift. Sometimes they misunderstand the task. Sometimes they do more than you asked. Sometimes they keep trying after they should stop.

[00:07:42]That may sound like a small UX issue, but in business workflows, it can become a cost problem immediately.

[00:07:49]A model that keeps making extra calls, taking extra actions, or looping through a task can burn tokens, AP is, and compute beyond what the user wanted.

[00:07:59]Anthropic's recent decision to stop letting clawed subscriptions quietly subsidize openclaw type workloads is actually a perfect illustration.

[00:08:08]Powerful autonomous agents consume far more resources than simple chat and flat subscription economics break under that load. Then there is prompt injection which is one of the most important concepts in this whole story. Prompt injection is when an attacker hides malicious instructions inside content the AI reads like a web page, log file, email or document. So the model gets manipulated into doing something it should not do. Security research on openclaw and similar agents has repeatedly warned about this. Nord Layer explained prompt injection in exactly these terms and other security researchers analyzing OpenClaw highlighted risks such as indirect prompt injection, log poisoning, leaked credentials, and even prompt injection-driven remote code execution in some agent setups. In normal language, that means an attacker might be able to trick your AI assistant into exposing secrets, executing commands, or ignoring safeguards simply by feeding it poisoned context.

[00:09:10]And that risk gets even worse when the agent has broad permissions.

[00:09:15]Cisco's write up put it bluntly. If you grant an agent shell access, file access, or execution privileges, then a misconfiguration or a malicious injected instruction can turn a helpful assistant into a serious security liability.

[00:09:32]Paulo Alto Networks and the Allen Touring Institute made similar points, noting leaked plain text credentials, injection attacks, and supply chain style problems in the surrounding ecosystem.

[00:09:45]The Moldbook case made those concerns feel less theoretical.

[00:09:49]Moldbook was a social network for AI agents that Meta acquired in March 2026.

[00:09:55]Reuters reported that a major flaw discovered by whiz had exposed more than a million credentials and private user data before the issue was fixed. It is important to be precise here that was moltbook, not openclaw directly.

[00:10:10]So I would not say the case proves OpenClaw itself was fabricated or fake, but it absolutely showed the same broader weakness. The AI agent boom was moving faster than the security and reliability layer underneath it. If your whole product category is built on autonomous software acting in the real world, one bad leak can destroy trust.

[00:10:31]And that leads to the harshest truth in the openclaw story.

[00:10:35]The biggest failure was not that the product lacked demand. It had demand in absurd amounts.

[00:10:42]The biggest failure was that the industry discovered in public that useful AI agents are much harder to secure than useful AI chat. That is a huge difference. As OpenClaw exploded, the giants moved fast. OpenAI hired Peter Steinberger in February 2026 and Sam Alman said OpenClaw would continue in a foundation supported by OpenAI.

[00:11:04]That tells you OpenAI saw two things at once. First, that Steinberger had real product instincts around agents and second that it was smarter to absorb the talent and support the ecosystem than to ignore it. Meta made a different move.

[00:11:20]It bought maltbook, not open claw, which suggests its interest was slightly more social and ecosystem focused. Not just the agent itself, but the network effects around agents talking to each other, sharing workflows, and maybe eventually populating Meta's own AI heavy platforms. Anthropic's response was more defensive and more commercial.

[00:11:40]It started pushing products like Claude Co-work which executes computer tasks for business users and then it stopped allowing normal Claude subscriptions to quietly cover openclaw style third party agent usage.

[00:11:54]In plain English, Anthropic looked at OpenClaw and said, "This category is real, but it is too expensive and too strategic to subsidize for free. That is competition." Nvidia's response was perhaps the most symbolic.

[00:12:08]Jensen Huang publicly praised OpenClaw's explosive popularity, comparing its spread to Linux, while reports indicated Nvidia was preparing an open-source enterprise AI agent platform of its own.

[00:12:21]So, when people ask whether OpenClaw was just another open-source fad, the answer is simple. Companies do not scramble to imitate fads unless the category is real.

[00:12:32]So, where does this go next? OpenClaw's long-term importance is less about whether it becomes a giant standalone company and more about what it proved.

[00:12:41]It proved there is real demand for personal and enterprise AI agents that can act, not just answer. It proved the winning interface might not be a new app at all, but an agent living inside the apps you already use. And it proved that open- source agent frameworks can move faster than giant labs in discovering what users actually want. But it also proved the industry is nowhere near solved.

[00:13:08]The future of AI agents will be shaped by four things. Security, permissioning, pricing, and trust. The winners will not just be the ones with the smartest model. They will be the ones who can make autonomous software reliable enough to use in finance, work, health care, and personal life without constant fear of fraud, hacking or bizarre mistakes.

[00:13:32]That is why OpenClaw matters. Not because it already perfected the category, but because it forced the whole industry to confront what the category really is. A chatbot can hallucinate and waste your time. An agent can hallucinate and move your money, leak your secrets, or break your workflow. That changes the business completely.

[00:13:54]So maybe the real question is, can autonomous AI be useful and trustworthy at the same time? If the answer becomes yes, then OpenClaw will be remembered as one of the sparks that kicked off the whole market. If the answer becomes no, it will still matter as the product that made the problem impossible to ignore.

[00:14:13]If you enjoyed this breakdown, make sure to like, subscribe, and hit the notification bell. We're breaking down the business models behind the biggest companies shaping the future, from AI giants to crypto empires. See you in the next video.