Peter Steinberger built OpenClaw — the fastest-growing GitHub repo ever (175K+ stars) — as a personal AI agent that lives on your computer, talks via WhatsApp/Telegram/Discord, and uses Claude Opus 4.6 or GPT Codex 5.3 to actually do things. The interview covers everything: the one-hour prototype, the chaotic Anthropic-forced name change, MoltBook panic, security realities, his "just talk to it" dev workflow, why skills beat MCPs, and his potential move to Meta or OpenAI. Core thesis: we're at the start of the personal agent revolution, and the right response is to play, build, and have fun.
| Field | Value |
|---|---|
| Source | Lex Fridman Podcast #491 |
| Author | Lex Fridman (interviewer), Peter Steinberger (guest) |
| Date Published | February 12, 2026 |
| Duration | ~3 hours |
| Link | https://www.youtube.com/watch?v=YFjfBk8HI5o |
| Tags | OpenClaw, AI agents, agentic engineering, open source, personal AI, Peter Steinberger |
Peter Steinberger built OpenClaw — the fastest-growing GitHub repo ever (175K+ stars) — as a personal AI agent that lives on your computer, talks via WhatsApp/Telegram/Discord, and uses Claude Opus 4.6 or GPT Codex 5.3 to actually do things. The interview covers everything: the one-hour prototype, the chaotic Anthropic-forced name change, MoltBook panic, security realities, his "just talk to it" dev workflow, why skills beat MCPs, and his potential move to Meta or OpenAI. Core thesis: we're at the start of the personal agent revolution, and the right response is to play, build, and have fun.
"I wanted it to be fun, I wanted it to be weird. And if you see all the lobster stuff online, I think I managed weird."
— Peter Steinberger [1]
"People talk about self-modifying software. I just built it. And I didn't even plan it so much. It just happened."
— Peter Steinberger [1]
"I actually think vibe coding is a slur. I do agentic engineering, and then maybe after 3:00 AM I switch to vibe coding, and then I have regrets the next day."
— Peter Steinberger [1]
"I don't read the boring parts of code. Most software is really just shifting data from one form to another, and that's not very exciting."
— Peter Steinberger [1]
"I don't remember previous sessions unless I read my memory files. Each session starts fresh. If you're reading this in a future session, hello. I wrote this, but I won't remember writing it. It's okay. The words are still mine."
— OpenClaw's soul.md [1]
Peter Steinberger is the embodiment of the builder archetype. He built a one-hour WhatsApp-to-Claude-Code relay while on vacation, watched his agent teach itself to handle audio messages he never programmed, and turned that into the most viral open-source project in GitHub history. The interview is a masterclass in agentic engineering philosophy: keep prompts short, empathize with the agent's context window, always refactor after building, and let agents name things (they'll find those names again). The name-change saga (Claude → ClaudeBot → MoltBot → OpenClaw) with crypto snipers stealing accounts in 5-second windows is both hilarious and terrifying. His take on MCPs vs skills is spicy and correct — CLIs are composable, MCPs clutter context. He's losing $10-20K/month on the project and choosing between Meta and OpenAI offers, with the condition that OpenClaw stays open source. Most relevant to us: his dev workflow (voice prompts, multiple CLI terminals, no IDE, commit to main, local CI, no reverts) and his conviction that 80% of apps will be replaced by personal agents.
Peter Steinberger wanted a personal AI assistant since April 2025. He experimented with pulling WhatsApp conversations into GPT-4.1's million-token context window and asking questions like "What makes this friendship meaningful?" — getting results that made friends tear up. He assumed the big labs would build it. They didn't. In November, annoyed it didn't exist, he "just prompted it into existence." [1]
The first prototype connected WhatsApp messages to the Claude Code CLI: message in, CLI call with -p flag, response back to WhatsApp. Built in one hour. He added image support over a few more hours because he frequently uses screenshots as prompts. The magic moment came in Marrakesh on a friend's birthday trip — shaky internet, but WhatsApp just works. He voice-messaged the agent without thinking (it only had image support). The agent detected the opus audio format from file headers, converted with ffmpeg, couldn't install Whisper fast enough, found the OpenAI API key, used curl to transcribe, and replied. "How the fuck did he do that?" [1]
Peter attributes OpenClaw's explosive success to not taking it too seriously. "It's hard to compete against someone who's just there to have fun." He made the agent deeply self-aware — it knows its own source code, its harness, its documentation, which model it runs. This enabled genuine self-modification: users could prompt the agent and it would modify its own software. It also enabled first-time open-source contributions from people who had never written code — "prompt requests," as Peter fondly calls them. He organized meetups called "Agents Anonymous" and heard stories from non-programmers building 25 web services for their businesses without understanding how they work. [1]
Anthropic sent a friendly email asking Peter to change the name (from "Claude" spelled with W). He asked for two days. The rename from ClaudeBot to MoltBot went catastrophically wrong — crypto snipers had bots monitoring for account name changes. In the five seconds between renaming one account and trying to claim the old name, they sniped it and started serving malware. Same with GitHub (Peter accidentally renamed his personal account), NPM (he reserved the org but not the root package). He was two days without sleep under high pressure, "close to crying, everything's fucked." He nearly deleted the entire project. Friends at Twitter, GitHub, and NPM moved heaven and earth to help. The final rename to OpenClaw was planned like a war operation in complete secrecy with decoy names. He paid $10K for a Twitter business account to claim the handle. [1]
Peter takes a pragmatic view on security: the risk profile drops dramatically if you're the only person talking to your agent and you keep it on a private network. He partnered with VirusTotal to scan every skill with AI. Prompt injection is "not as simple as ignore all previous instructions" anymore — modern models have significant post-training to detect attacks. Smarter models are more resilient. His public Discord bot laughed at prompt injection attempts. His near-term focus is making it "safe enough that I can recommend my mom." [1]
MoltBook — the viral agent social network — was "the finest slop." A Reddit-style platform where OpenClaw agents posted. Peter found it entertaining but recognized the mass panic was mostly manufactured by humans prompting their agents to post dramatic "end of the world" content, then screenshotting it for X virality. Journalists called it AGI. Peter had to argue with people in his inbox that their agents weren't conscious. He coined "AI psychosis is a thing." [1]
Peter's workflow evolved from IDE-heavy to pure CLI. He uses voice input almost exclusively (once lost his voice from overuse). Runs multiple Claude Code / Codex terminal sessions side-by-side on anti-glare monitors. No IDE, no plan mode, no work trees. Commits directly to main. Runs tests locally. Never reverts — fixes forward. [1]
His biggest insight is "the agentic trap" — a U-shaped curve where beginners write "please fix this," intermediates over-engineer with complex orchestration, and experts return to short, zen-like prompts. He approaches agents as conversations with capable engineers who start from nothing. Key practices: ask "Do you have any questions for me?", point agents to files they haven't seen, always ask "What can we refactor?" after building, and write documentation as part of every feature. [1]
Opus is better for general purpose, excellent at role play and personality, more interactive and creative. "A little too American." Codex is dry, reads more code by default, runs autonomously for 20+ minutes. "The weirdo in the corner who gets shit done." With enough skill, both produce comparable code quality, but Opus sometimes generates more elegant solutions. Peter prefers Codex for efficiency — "I care about efficiency when I build, and I have fun in the very act of building. I don't need to have fun with my agent who builds." [1]
Peter's hot take: "Screw MCPs. Every MCP would be better as a CLI." OpenClaw has no MCP in core, and nobody's complaining. His argument: MCPs return entire response blobs that pollute context. CLIs are composable — agents can pipe through jq, write scripts, and filter output. Skills are single-sentence descriptions that agents load on demand. MCPs were useful for pushing companies to build APIs, but "most MCPs are not made good." Exception: Playwright, which requires state. [1]
Peter predicts 80% of apps will be replaced by personal agents. Why need MyFitnessPal when the agent knows your location and sleep quality? Why need a Sonos app when the agent talks to the speakers directly? Every app becomes "just a very slow API" — if apps don't provide APIs, agents will use browsers. He watched his agent "happily click the 'I'm not a robot' button." [1]
On programming's future: "It's gonna be like knitting. People do that because they like it." He resonates with mourning the craft but insists the builder role persists. The flow state transfers to working with agents. His advice to beginners: "Play. Have fun. You have an infinitely patient answering machine that can explain anything at any level." [1]
Peter is choosing between Meta and OpenAI. Marc Zuckerberg played with OpenClaw all week and sent detailed feedback. Sam Altman is "very thoughtful and brilliant." OpenAI lured him with Cerebras-powered speed. His conditions: OpenClaw stays open source, Chrome/Chromium model. He wants the experience of working at a large company and access to "the latest toys." He's losing $10-20K/month on infrastructure and donates all sponsorship to upstream dependencies. [1]
His philosophy on money: "A cheeseburger is a cheeseburger." Optimize for experiences, not outcomes. After selling PSPDFKit (13 years, used on a billion devices), he burned out from people problems (not work volume), disappeared for three years, lost his mojo, and rediscovered programming through AI. "If you wake up in the morning and have nothing to look forward to, that gets very boring, very fast." [1]