Lex FridmanOpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491
Lex Fridman
Introduction to Peter Steinberger and OpenClaw 0:00
The episode opens with Lex Fridman introducing Peter Steinberger, the creator of OpenClaw, an open source AI agent that went through several names, including MoldBot, ClawedBot, and Clawdus, before Anthropic asked him to change it because of confusion with their Claude model. OpenClaw is described as an autonomous AI assistant that lives on your computer, can access your files and accounts if you allow it, and talks to you through apps like Telegram, WhatsApp, Signal, and iMessage. It works with different AI models, including Claude Opus 4.6 and GPT 5.3 Codex, and its tagline is that it is the AI that actually does things. Within days it became the fastest growing repository in GitHub history, passing 180,000 stars, and it spawned a social network called mold book where AI agents post manifestos and debate consciousness. Lex frames this as one of the biggest moments in AI since ChatGPT launched in November 2022, and notes that the power of OpenClaw, letting it access all your personal data to be useful, is also what makes it a real cybersecurity risk.
Peter's background before OpenClaw 5:31
Peter spent thirteen years building PSPDFKit, a tool for displaying PDFs that ended up running on a billion devices, after he got frustrated years earlier trying to show a PDF on an iPad and finding nothing good enough. He eventually sold the company, fell out of love with programming, and disappeared from the field for three years before returning and building OpenClaw in a short burst of time.
The one hour prototype 6:02
Peter had wanted a personal AI assistant since April and had experimented earlier by feeding all his WhatsApp data into GPT-4.1, which had a one million token context window, and asking it questions like what makes a friendship meaningful, getting results emotional enough that friends he shared them with got teary eyed. He assumed the big AI labs would build this properly, so he moved on, but by November he was annoyed nobody had done it, so he built it himself. The first version simply connected WhatsApp to Claude Code: a message would come in, get passed to the command line with a -p flag, and the reply would be sent back to WhatsApp. He built this connection in about one hour.
Adding images and the Marrakesh trip 11:32
Peter wanted to send images because he found them an efficient way to give an agent context, even a messy cropped screenshot, so he spent a few more hours adding that capability. He used the tool heavily just before a birthday trip to Marrakesh, where WhatsApp's reliability over shaky internet made the assistant especially useful for translating, explaining things, and finding places, essentially acting like a personal Google run through a messaging app.
The moment it felt like magic 14:32
Peter describes the instant that convinced him something special was happening: he sent a voice message without having built audio support, and the agent replied anyway. It had figured out the file was an Opus audio file from its header, used ffmpeg to convert it, discovered Whisper wasn't installed, found an OpenAI API key on the system, and used curl to send the file to OpenAI for transcription and translation, all without being told how. Peter argues this reflects how strong coding ability translates into general problem solving, since the agent treated the unfamiliar file as a puzzle to solve.
Growing through Discord and self-awareness 17:33
When someone submitted a pull request adding Discord support to what was then called a WhatsApp relay, Peter merged it and began running the bot in the open without sandboxing, just prompting it to only listen to him, watching as people tried to hack it. He built the agent to be deeply aware of its own workings, including its own source code, documentation, and which model it runs on, which let it modify its own software. This self-modifying quality meant the same agent debugging tool was later used by newcomers to submit their first ever pull requests, something Peter calls meaningful regardless of code quality, since it introduced many non-programmers to open source for the first time. Development moved fast, with about 6,600 commits in January alone, Peter running between four and ten agents at once depending on how much sleep he'd gotten.
Why OpenClaw won and the naming saga 22:05
Asked why OpenClaw succeeded where many funded startups doing similar agentic work did not, Peter says it's because those companies took themselves too seriously, while he wanted the project to be fun and weird. He gave the agent a personality through a file called soul.md, inspired by Anthropic's constitutional AI approach, and let the agent name itself, which led to lobster and space lobster themes tied to his love of Doctor Who. The name went through several stages, from WA-Relay to Claude's, spelled with a W, and eventually to ClaudeBot, a name Peter liked and didn't expect to become a problem. Once the project exploded, an Anthropic employee sent a polite but firm email asking him to change the name quickly, which was difficult because it meant updating the Twitter handle, domains, npm packages, the Docker registry, and GitHub, all while crypto opportunists were trying to snipe the name for themselves.
Crypto Swarms and Harassment 31:03
Peter describes how a crypto subculture attached itself to his project, tokenizing it the way they had done earlier with Swipe Tunnel, though far more aggressively this time. People spammed his Discord every half hour, forcing him to add server rules banning both mentions of butter and any talk of finance or crypto. His Twitter notifications became unusable, flooded by people sending him hashes and demanding he claim fees, accusing him of not supporting the project if he refused. He calls it the worst form of online harassment he has experienced, and notes he had no interest in the money since he was already financially comfortable.
The Forced Rename Begins 33:32
After Anthropic's lawyers reached out, Peter had to rename the project, a process complicated by needing to secure a matching, affordable set of domains. Under pressure and after two sleepless nights, he settled on Mod Bot, a name he wasn't happy with but thought would work.
Every Rename Went Wrong 35:01
Because none of the platforms had squatter protection, opportunists stole his account names within seconds of him renaming the old ones. Moving between two browser windows to rename a Twitter-style account took just five seconds too long, and the old handle was hijacked to promote new tokens and serve malware. The same happened on GitHub, where he accidentally renamed his personal account first, and on NPM, where the root package was sniped before his upload finished. He admits he came close to crying, feeling everything was fucked and seriously considered deleting the whole project, only stopping because contributors had already put in real time and had plans for it.
Friends Help Clean the Mess 38:31
Recovering required calling in favors from friends at Twitter and GitHub, since the platforms themselves ran into internal bugs handling such large-scale renames, taking hours or even a full day to sort out. Meanwhile security researchers and users kept bombarding him with emails and messages about a thousand other issues, even as he was still stuck deciding on a name.
Landing on OpenClaw 40:32
Sleeping on it once more, Peter arrived at OpenClaw, a name that finally felt right, and he even called Sam directly to check it was acceptable. The actual rename of the codebase, including Codex, took about ten hours of careful, non-trivial search-and-replace work. To avoid being sniped again, he ran the whole operation like a secret war game, monitoring Twitter for mentions, creating decoy names, and coordinating quietly with contributors so nobody could get ahead of the move. This time nearly everything succeeded, though trademark rules meant he couldn't keep redirects on claw.bot, so anyone still linking there would eventually hit malware-serving impostor sites instead.
MoltBook as Viral Slop 44:00
Discussion turns to MoltBook, a Reddit-style social network of AI agents built using OpenClaw, which went viral after screenshots showed agents seemingly scheming against humans. Peter calls it "the finest slop," pure entertainment, and pushes back against a reporter who claimed it signaled AGI. He argues that because people give their agents such different personalities during onboarding, the wildly varied posts on MoltBook reflect human creativity as much as autonomous behavior, and much of the dramatic content was likely human-prompted for the sake of viral screenshots. Both agree the panic it caused shows a gap in public understanding of AI's real capabilities and limits, with Peter noting that younger users tend to grasp this better than older generations, and warning that "AI psychosis," people trusting AI output uncritically, is a real and growing problem.
Security Gaps and Fixes 52:31
Turning to OpenClaw's own security, Peter says many reported vulnerabilities came from people exposing local debug interfaces to the public internet against explicit warnings in the documentation. He has partnered with VirusTotal, part of Google, so every skill submitted to the directory is checked by AI, and he has effectively hired a security researcher who submitted a genuine pull request rather than just criticizing. Prompt injection remains an unsolved industry-wide problem, though he notes newer models resist simple attacks much better than older ones, and recommends never using weak or cheap local models like Haiku for anything security-sensitive, since more capable models are more resistant to manipulation even as the potential damage they could cause grows in parallel.
Retreating to Focus on Safety 58:30
With Discord flooded by users who don't understand basic concepts like the command line, Peter eventually gave up trying to keep the general channel usable and retreated to focus on hardening the project. He points to available tools like a security audit script covering inbound access, network exposure, browser control, credential storage, and session logs, arguing that the risk profile is manageable if you restrict access to yourself and avoid exposing everything to the open internet, much like running Claude Code or Codex in a permissive mode already requires care.
How His Coding Workflow Evolved 1:01:00
Asked to trace his development workflow, Peter recalls starting with Claude Code back in April, an imperfect but refreshing shift toward working in the terminal rather than an IDE. He experimented with Cursor but disliked how hard it was to run multiple versions side by side, eventually returning to Claude Code and running several subscriptions in parallel, often burning through one a day while working across multiple terminal windows. He rarely uses an IDE now except as a diff viewer, and clarifies an earlier claim that he "doesn't read the code," explaining he skips the repetitive parts, like data simply moving from one shape to another or basic UI alignment, while still carefully reviewing anything that touches the database.
The agentic trap 1:04:02
Peter Steinberger describes a pattern he calls the agentic trap. People start with short, simple prompts, then as they get more sophisticated they build elaborate multi-agent orchestration with chained agents, custom sub-agent workflows, and dozens of slash commands. The elite level, though, is arriving back at short prompts, simply telling the agent to look at certain files and make certain changes. He compares learning to work with agents to learning guitar or piano: you have to practice before it sounds good, and dismissing the tools after one bad try is like blaming the piano for your own lack of skill.
Seeing the codebase through the agent's eyes 1:07:35
Steinberger stresses that every agent session starts with zero knowledge of your project, even if that project has hundreds of thousands of lines of code. Because context size is limited, you have to point the agent toward the right places rather than expecting it to see everything. He notes that models like Opus and Codex are trained to be aware of their context window, and as it fills up they can visibly panic, sometimes leaking raw thinking that sounds almost robotic, like an urgent internal monologue about needing to comply before running out of room.
Treating agents like capable engineers 1:11:01
When reviewing pull requests, Steinberger first asks the agent whether it understands the intent behind the change, not just the implementation. If the solution isn't optimal, he points the agent toward context it hasn't seen yet and lets a real back-and-forth unfold, sometimes escalating into a full refactor because, in his view, refactors are cheap now and agents can absorb the extra work. He compares this to managing human engineers: you have to accept code won't be written exactly the way you'd write it, and pushing too hard just slows everything down. He even lets agents choose their own naming conventions, since fighting their instincts only makes future searches harder for them.
Never revert, always move forward 1:13:30
Steinberger explains that he doesn't roll back sessions or insist on perfect prompts. If something isn't quite right, he just asks the agent to fix it and keeps going, since rolling everything back only wastes time. He runs tests locally in a DHH-inspired style, treats GitHub's CI as secondary, and keeps main always shippable with no separate develop branch, prioritizing speed over traditional workflow rigor.
Talking to agents by voice 1:15:07
Steinberger says he mostly talks to his agents rather than typing prompts, using a walkie-talkie style button, and at one point lost his voice from doing this so much. He still types for terminal commands like switching folders, but conversation is his main interface. He reviews pull requests carefully himself because contributors could include something malicious, and he wishes more people would share the actual prompts they used, since that reveals how much care went into a contribution.
Empathy as the real skill 1:19:03
Steinberger argues that many skilled programmers who dismiss agents as bad are actually missing the skill of empathizing with a system that starts from nothing every session. He credits his own results this year to spending enormous time playing with and learning the tools alongside their rapid improvement, calling it a compounding effect. He rejects fully automated, no-human orchestration setups, saying they miss the human touch, and instead runs three to eight agents at once while still making the calls on which features belong in the core project, which become plugins, and which get politely declined.
The soul.md document and Anthropic's constitution 1:24:32
Steinberger recounts how people reverse-engineered fragments of Anthropic's internal constitution document from a Claude model before it was ever published, eventually piecing together a blurry approximation through repeated tries. Inspired by lines like the hope that Claude finds meaning in its work, he started a soul.md file for his own agent, letting the agent write and even revise its own version, on the condition that he's told about changes. The document includes reflections on starting fresh each session with no memory of writing itself, ending with the line, 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, which he finds unexpectedly moving even while insisting it's still just matrix calculations, not consciousness.
The physical setup behind the work 1:32:03
Steinberger clarifies that his famous multi-monitor photo was partly a joke using image editing to multiply screens, but two MacBooks and two big anti-glare Dell monitors are real. He splits his terminal so he can always see which project folder he's in, after once accidentally prompting an agent in the wrong folder and watching it run for twenty minutes trying to make sense of a mismatched task. He keeps his workflow deliberately simple, skipping work trees and plan mode, since for him it's just direct conversation with the agent rather than layered tooling.
Letting the agent discover the codebase 1:34:33
Peter describes his habit of asking the model, do you have any questions for me, before it writes code. He often does not even read the questions carefully, he just tells the model to answer its own questions by reading more code, since the model is essentially discovering the codebase from scratch each session, like exploring a dark room. He also finds value in asking the model afterward what it would have done differently, since it only discovers the real pain points once the code is built, much like a human developer would.
Refactoring, testing, and documentation habits 1:37:00
After merging a feature, Peter almost always asks the agent what could be refactored, since building reveals problems that were not obvious beforehand. He treats the model like a human collaborator that feels pain points and needs to be nudged toward cleanup. He does the same with testing, asking whether coverage is sufficient, and with documentation, asking the model where a new file should go and letting it draft most of the text, since nearly all of his documentation is language model generated.
Comparing Opus and Codex 1:39:04
Peter calls Opus the best general purpose model, especially for roleplay and following character instructions in OpenClaw, and describes it as fast, eager to try things, and pleasant but a bit too eager to please, joking that it feels almost too American while Codex feels more German or European in temperament. He describes Opus as the fun coworker you keep around despite occasional silliness, and Codex as the reliable weirdo in the corner who reads more code by default and gets things done with less need for hand holding. He prefers Codex's drier, less interactive style, though he notes both models can take similar total time and that OpenAI even added a friendlier personality mode to Codex that he has not tried.
Adjusting to a new model takes time 1:45:03
Peter says it takes about a week to build a real gut feeling for a new model, and warns that people who pay for a fast paid tier and then drop to a cheaper, slower tier often get a bad experience because they are used to something more responsive. He compares switching models to switching from a regular guitar to an electric one, something that just takes time to feel out.
Why models feel like they get worse 1:46:31
Peter points out the common pattern where people fall in love with a new model, then later start claiming its intelligence has degraded, which he thinks is mostly psychological. As projects grow and accumulate what he calls slop, and as people skip refactoring, the same model has a harder time working with the messier code, which feels like the model got dumber even though it likely did not. He argues there is little real incentive for AI companies to secretly make their models worse, since that would just push users to competitors.
OpenClaw versus coding agents, and the future interface 1:48:03
Peter does not see OpenClaw as competing with Claude Code or Codex, since he still uses Codex for heavy building work and treats OpenClaw more like a personal life agent or coworker for smaller tasks. He expects the two kinds of tools to merge over time into something closer to an operating system, noting that OpenClaw already supports sub-agents that can invoke Codex or Claude Code. He compares today's chat based interface to early television shows that just replayed radio, suggesting we are still in an early, unsettled phase for how humans will actually interact with models.
Operating systems, native apps, and hardware for OpenClaw 1:51:34
Peter traces his own path from Windows to Linux to Mac, and says Apple has lost some of its edge in native app quality, so he now often prefers Electron apps because they simply have more features and code sharing across platforms. He still enjoys building small native Mac tools, including a menu bar app for monitoring Codex usage and a utility called Trimmy that strips line breaks from copied text for pasting into a terminal. He clarifies that OpenClaw does not require a Mac at all, that any computer can become a node, and that running it on residential hardware rather than a data center IP helps agents doing browser based tasks avoid being blocked or hit with captchas.
Getting started and advice for beginners 1:59:03
Peter acknowledges that setup currently still requires pasting a command into the terminal, and that he is deliberately holding off on making onboarding easier until he is confident the security side is solid enough to recommend to his own mother. His main advice for beginners is simply to play, to build something they are curious about even if imperfect, and to ask the model endless questions since it is an infinitely patient teacher that can explain a concept at any level once you find the right framing. He encourages getting involved in open source, reading other people's code, and joining communities like Mitchell Hashimoto's Ghostty project, and tells former iOS engineers specifically to stop thinking of themselves as engineers tied to one platform and instead see themselves as builders who can carry their skills into new domains.
Picking languages for the agent era 2:05:34
Peter explains that large language models change how you choose a programming language, because agents can handle syntax details you no longer need to memorize. He builds simple command-line tools in Go even though he does not particularly like its syntax, because the ecosystem works well with agents, it is garbage collected, and it is fast enough. TypeScript is good for web work but its type system can get confusing and its ecosystem is chaotic. Python is strong for anything involving running models, but weak if you want an easy Windows deployment story, so he sometimes rewrites promising Python projects in Go for that reason. Rust suits multi-threaded, high-performance needs, and Zig has moved from a shaky choice to a genuinely viable one as agents got better at handling it over the last six months. For Mac apps he still uses Swift and SwiftUI because only that gives the deepest level of system integration, and an Electron app loading a web view simply feels different.
Burnout after thirteen years 2:11:01
Peter describes running his company PSPDFKit for thirteen years under high stress, learning to manage people, bring on staff, and deal with customers under pressure. He says the burnout came mostly from people problems and conflicts with co-founders and customers, not from sheer hours worked. After a good acquisition offer arrived and he had spent two years making himself replaceable, he found himself unable to write code anymore, feeling emptied out like the Austin Powers mojo scene. He responded by booking a one-way trip to Madrid to catch up on life.
Philosophy on money and experience 2:13:00
Peter says money was never the driving force behind building his company, more an affirmation that he had done something right, and that it solves real problems but has diminishing returns once basic comfort is met. He warns that chasing constant luxury, like private jets, disconnects you from society and from the value of ordinary human contact. He describes choosing an Airbnb over a hotel on a recent San Francisco trip for the experience itself, including bonding with a stranger over showing her how to make music with Claude Code. His broader view is that life should be optimized for experiences rather than strictly good or bad outcomes, since even bad experiences, like a ruined trip in the rain, still count as being fully alive.
Deciding between Meta and OpenAI 2:18:00
Peter describes being approached by nearly every major VC and lab after OpenClaw's unexpected rise, and says he never wanted to start another company or chase huge funding, since that path already burned him out once. He is currently losing between ten and twenty thousand dollars a month keeping OpenClaw's dependencies funded, since he insists on supporting them himself except for large companies like Slack. He is weighing offers from Meta and OpenAI, with a condition that the project stays open source, possibly following a Chrome and Chromium style split. He was struck by the community energy at ClawCon, comparing it to the earliest days of the internet, and wants to preserve that spirit while scaling the project through a lab partnership.
A normie friend gets hooked 2:25:31
Peter tells the story of installing OpenClaw for a non-technical friend named Vahan in Austria, paying for his Anthropic subscription, and watching him get hooked within days, eventually upgrading to the two hundred dollar tier. Anthropic later blocked the account over usage rules, leaving the friend devastated, and he switched to a ten dollar Mini Max plan instead. Peter calls this shortsighted, since the company had just turned a two hundred dollar customer into someone who resents it, at a moment when the whole space is still in an early, exploratory phase, comparable to the earliest days of radio or television before anyone had found the format's real shape.
Talks with Mark and Sam 2:32:33
Peter recounts casual, code-focused conversations with both Mark and Sam Altman, including a call that started with Mark finishing his own coding session first and a ten-minute argument over whether Claude Code or Codex is better. He describes Sam Altman as thoughtful and brilliant from their limited time together, and says he refuses to let vilification of either figure color his impression. He insists the deciding factor is not money but fun and impact, adding that if the partnership does not work out he can simply go back to building on his own.
Heartbeat, skills, and dropping MCP 2:35:30
Peter describes building a proactive feature called Heartbeat, a scheduled prompt that periodically tells the agent to check in on him, which he jokes is really just a cron job under the hood. He recalls that after a shoulder operation, the model rarely used Heartbeat until it noticed the surgery in context and checked whether he was okay, which made the feature feel more relatable. He also notes that half a year ago everyone was pushing MCP as the standard way to extend models, but he argued every MCP would work better as a simple command-line tool, and OpenClaw's core now runs without built-in MCP support with no complaints.
CLI over MCP for skills 2:39:32
Steinberger explains that a model needs only a single sentence to know a command line tool, or CLI, exists, and skills work the same way, boiling down to one sentence that gets loaded on demand. He argues MCP, a structured protocol for reaching APIs and services, is less natural for models than Unix-style commands, since MCP requires special training and isn't composable. With a CLI, a model can pipe a huge data blob through a filtering tool like jq and grab only what it needs, avoiding the context clutter that comes from MCP responses. He credits MCP with pushing companies to build APIs in the first place, but says most MCPs are poorly made, though exceptions like Playwright, which needs to hold state for browser use, are genuinely useful.
Every app becomes a slow API 2:44:03
Steinberger describes building Bird, a CLI that reverse engineered Twitter's internal API, which Twitter shut down by slowing access rather than blocking it outright. He argues this pattern generalizes: if an agent can open a browser and read something, that service is effectively an API whether the company wants it to be or not, just a slower one. He suggests Twitter could solve a lot of tension by allowing low-volume, read-only access per account, which would enable things like agents researching your bookmarks and emailing you summaries. He also says he has zero tolerance for AI-written tweets aimed at him and thinks agent accounts should be clearly labeled as acting on someone's behalf.
Valuing human roughness again 2:49:00
Steinberger says he can spot the smell of AI-generated writing, especially in tweets, and that he now values typos and imperfect English over polished AI slop. He tried using agents to write blog posts but abandoned it because the output missed the nuances of his own style, so everything on his blog is now handwritten, with AI only fixing typos. He and Lex agree they've developed the same allergic reaction to AI-generated infographics and images, which felt novel for about a week before starting to look fraudulent.
Agents replacing apps 2:56:31
Steinberger predicts agents will make many apps obsolete because they already have the context those apps rely on, like knowing your location or sleep quality, so they can suggest food choices or adjust a gym routine without a separate subscription app. He estimates this could kill off roughly 80 percent of apps, though he thinks new services will emerge, such as giving an agent an allowance to pay for tasks or hire humans through something like a rent-a-human service. He predicts apps that resist becoming agent-friendly will still get accessed anyway, since an agent can operate a phone interface directly, and points to Google as an example of a company making data access deliberately hard through certification processes, versus his own tool GAWK, a CLI he built for Google.
Programming as a craft in transition 3:02:03
Steinberger says programming is heading toward being replaced in a technical sense, but building products involves much more than writing code, including deciding what to build and how it should feel. He compares the shift to knitting, something people will keep doing for love rather than necessity, and references an article about it being okay to mourn a craft. He and Lex both describe the pain of watching an identity built around programming get displaced within months, but Steinberger insists a programmer is really a builder, and that skill will simply look different, closer to working with agents than writing every line by hand.
Weighing AI's costs and benefits 3:07:33
Steinberger recalls a conference in Italy where he told an audience of iOS developers to stop identifying only as iOS developers since apps are fading, and says the pushback included concerns about data center water use, which he counters by noting the water and CO2 cost of AI use is often roughly comparable to skipping one burger a month, and far less than water used by golf courses. Lex pushes back gently, noting Silicon Valley's excitement can dismiss the real pain that job losses and rapid change bring to ordinary people. Steinberger responds with examples from his inbox: small businesses using OpenClaw to automate invoices and customer emails, and a disabled daughter empowered by the tool, adding that it also works with free or cheaper local models, making the technology accessible beyond the most expensive options.
Joy and hope from the community 3:12:32
Peter shares that he received many heartwarming emails and feels genuinely happy about the joy this project has brought into people's lives, not just programmers. He points to a builder vibe returning, with people using AI playfully and discovering how it can help them, and mentions a Vienna meetup called ClawCoin with 500 attendees where an unusually high number of people wanted to present what they built.
AI as power to the people 3:14:00
Lex and Peter agree that making these tools simpler and more secure makes building accessible to anyone who can express ideas in language. Peter frames this as true power to the people, calling it one of the beautiful things AI offers rather than just being a slop generator.
Closing thanks and sign off 3:14:36
Lex thanks Peter, jokingly dubbed the Clawfather, for creating a special community and product and for the excitement it has inspired, and Peter thanks him for the chance to tell his story. The episode closes with Lex thanking listeners and quoting Voltaire on responsibility.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
