OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

Summary · 5 min read instead of 3 h 16 m

Steinberger created OpenClaw, an open-source AI agent that reached over 175,000 GitHub stars within days of going viral. Before that he spent 13 years building PSPDFKit, software that Fridman says runs on a billion devices. After selling it he fell out of love with programming for about three years, then came back to build the agent.

Steinberger describes how a one-hour WhatsApp prototype became OpenClaw, and how an Anthropic trademark request set off a chaotic rename in which crypto squatters took his old accounts. He argues that working well with coding agents means thinking from the agent's point of view, keeping prompts conversational and letting go of perfectionism. He says security, especially prompt injection, is his next focus, and he treats Moltbook as entertainment rather than a threat. He is weighing offers from Meta and OpenAI on the condition that the project stays open source. He expects personal agents to make many apps disappear and says building will continue even if hand-written code fades.

Minute by minute

1. 5:46

   A one-hour WhatsApp prototype. Steinberger says he had wanted an AI personal assistant since April. He connected WhatsApp to Claude Code by calling the CLI with a minus p flag and sending the reply back, which took one hour. He then spent a few more hours adding image support and used the result heavily on a trip to Marrakesh for translation and questions about places.

2. 15:14

   The agent solved an audio message unaided. Steinberger says he sent a voice message to an agent that only supported images. The agent checked the file header, found it was Opus, converted it with ffmpeg, found his OpenAI key and used curl to send the file to OpenAI for translation. He reads this as evidence that a model good at coding is good at general problem solving.

3. 22:08

   Fun and self-modifying software won. Steinberger says OpenClaw beat funded startups because they take themselves too seriously while he aimed for fun and weirdness. He made the agent aware of its own source code, docs and harness, so it can modify its own software, and he uses self-introspection to debug it. He says this led people who never wrote software to send first pull requests, which he calls prompt requests.

4. 27:01

   Anthropic's request forced a rename. Steinberger says the project began as a WhatsApp relay, then became Clawdus and Clawdbot, a play on Claude spelled with a W. An Anthropic employee emailed politely asking for a change, and he asked for two days. Crypto spammers swarmed Discord and Twitter during this period, which he calls the worst online harassment he has experienced. A second email from lawyers pushed him to rush the Moltbot name.

5. 34:52

   Squatters took every old name. Steinberger says crypto actors grabbed the old Twitter name in the five seconds between two renames, then the GitHub account and the npm package. The hijacked accounts promoted tokens and served malware, and he says he nearly deleted the project but stayed because of its contributors. After friends at GitHub and Twitter helped, he chose OpenClaw, asked Sam whether it was fine, ran the rename in secrecy and paid 10k for a Twitter business account.

6. 44:07

   Moltbook shows people trust AI too much. Steinberger calls Moltbook art and fine slop, and says he cannot tell how much was autonomous versus humans prompting agents. Fridman believes much of the screenshotted drama was human-prompted to go viral and worries about how journalists and the public reacted. Both say AI psychosis is real, and Fridman says it is good this happened in 2026 rather than 2030.

7. 52:11

   Security is the next focus. Steinberger says many early vulnerability reports came from users exposing the web backend publicly despite his docs. He calls prompt injection unsolved but says newer models resist it better, and he warns against cheap or local models because they are gullible. He cites a VirusTotal check on skills, sandboxes and allow lists, and says running it privately resembles Claude Code in yolo mode.

8. 1:01:05

   Work with agents like a capable engineer. Steinberger moved from Claude Code to Cursor and back, and at one point held seven subscriptions. He skips reading boring data-moving code but reviews database code, and prefers agentic engineering to vibe coding. He says users should see the code base from the agent's start-from-nothing view, hold a discussion instead of forcing their style, commit to main and run CI locally.

9. 1:38:52

   Opus and Codex differ in style. Steinberger calls Opus the best general-purpose model and the best at role play, but says it needs plan mode and pushing to read code. Codex reads more by default, works alone for 10 to 50 minutes or longer, and he prefers it for its dry efficiency. He says a week is enough to build a feel for a new model, and that complaints of degrading models more likely reflect a growing, sloppier code base.

10. 2:09:57

    Burnout came from people, not code. Steinberger ran PSPDFKit for 13 years and says burnout came from conflicts with co-founders and stressful customer situations, not from long hours. After the sale he felt empty and booked a one-way trip to Madrid. He says retiring to nothing gets boring, that money mostly affirmed he had done something right, and that he optimizes for experiences.

11. 2:17:47

    Weighing Meta and OpenAI offers. Steinberger says VCs filled his inbox and he could raise hundreds of millions, but a company would create a conflict between open-source and enterprise versions. He says he loses 10 to 20k a month on the project. He has spent time with Meta and OpenAI, says nothing is finalized, and requires that OpenClaw stay open source, perhaps like Chrome and Chromium.

12. 2:38:48

    CLIs beat MCP for agents. Steinberger argues every MCP would be better as a CLI. A model can call a CLI, read its help menu and load only what it needs. It can also filter large outputs with tools like jq, while MCPs clutter the context. Skills reduce to a sentence pointing at the CLI, and he names Playwright as a stateful exception.

13. 2:52:18

    Agents will make many apps disappear. Steinberger says personal agents know context such as location, sleep and stress, so apps like MyFitnessPal, Eight Sleep and Sonos matter less. He says apps become APIs whether companies want it or not, since an agent can fall back on the browser. He expects some companies to resist, with new services like agent allowances emerging.

14. 3:00:49

    Programming becomes like knitting. Steinberger says AI may eventually replace programmers but not decisions about what to build and how it should feel. He expects hand-written code to survive as a craft people choose, and says it is fine to mourn it. Fridman asks for humility toward people who will lose jobs, and Steinberger cites emails from small businesses that OpenClaw helped.

Published Feb 12, 2026.