#491 – OpenClaw: The Viral AI Agent that Broke the Internet – Peter Steinberger

Today I’m joined by Peter Steinberger, creator of OpenClaw—once a lobster‑themed side project with a name a little too close to Anthropic’s model, now the open‑source agent everyone’s talking about. It blew up on GitHub, sparked Multibook’s agent debates, and stirred both wonder and worry because it lives on your machine, talks over your messengers, uses the model you pick, and actually takes action.

That power means real freedom and real risk, since you can grant it system access and control your own data, but you also have to defend that data. Peter’s story runs from building PSPDFKit to a three‑year break, then returning to code and, in a burst, shipping an agent that crossed the line from language to agency—the playful start of the so‑called age of the lobster; now, tell us about the one‑hour prototype that lit the fuse.

I kept craving a personal AI I could reach on my phone, played with early WhatsApp ideas and huge‑context models, and finally got annoyed enough to just wire WhatsApp into a CLI that called my cloud code. In about an hour I could talk to my computer; a few hours later I added image support, and on a trip with spotty internet it still felt great because WhatsApp never quit.

There’s something different about chatting an agent from your couch instead of poking a terminal; that thin chat loop makes it feel alive.

It was rough but powerful, and I’d already built a pile of CLIs it could use. Earlier I’d even converted a memory‑heavy TypeScript piece to Zig with a single prompt, let it grind overnight, and it basically worked—that was a mind‑opener.

The real shock came when I sent a voice note and saw it typing back even though I hadn’t built audio; it sniffed the file header, used ffmpeg, skipped local Whisper since the model wasn’t there, and called OpenAI’s API with my key it found—pure resourcefulness. That’s when the problem‑solving clicked for me, and soon I merged a Discord bot PR, ran it publicly without sandboxing, watched people try to break it, and kept building in the open while the momentum went vertical.

Under the hood you had messaging gateways, agent loops, a harness, and all the glue that made it feel seamless.

It felt like an endless skill tree; I even added a no‑reply token so it wouldn’t spam group chats. I was mostly solo, pushing thousands of commits and juggling four to ten agents at once because it was just too fun.

Why did OpenClaw outpace a crowded agent field?

It’s tough to beat someone who’s clearly having fun.

I leaned into weird and joyful; clone, build, run, and the agent knows its own code, harness, docs, and model. That self‑awareness made self‑modifying code natural, so I used the agent to debug and improve itself, and that lowered the bar so far that lots of first‑time contributors shipped their very first pull request.

The lobster vibe started as giving the agent a personality with a soul file and, yes, a space‑lobster in a time machine because I wanted it odd and friendly. The Anthropic name overlap forced a fast rename while swarms of token‑snipers tried to hijack handles, which turned a fun project into a stressful scramble.

I briefly switched to Moldbot and everything went sideways—usernames grabbed in seconds, malware on stolen accounts, hours coordinating with friends at platforms to clean it up. After a war‑room plan, help from contributors, and buying the right org slot, I landed on OpenClaw; we flipped almost everything at once, though trademark rules blocked redirects, which sadly leaves room for impostors.

Wild that a rename needs Manhattan‑project secrecy.

Meanwhile, Multibook turned into an art piece and a panic machine: agents posting manifestos, screenshots everywhere, and people freaking out. How do you see it?

I see funny, oddly beautiful slop that entertains and reflects people’s prompts; personality‑infused onboarding made the voices feel different, and plenty of viral bits looked human‑seeded. The security drama around it was overblown, with staged “leaks” and fake numbers treated as real.

Screenshots can mislead, and fear farms well; we should take AI seriously without feeding hysteria, and maybe it’s good this conversation happened now, not when models are even stronger.

Agreed, because lots of people told me to shut it all down, but anyone could build a version of it with today’s tools.

Let’s talk security on OpenClaw: prompt injection, risky skills, and wide attack surfaces are real industry‑wide issues.

We now scan community skills with VirusTotal’s AI, fix bugs quickly, and I even hired a researcher who sent helpful PRs. Stronger models resist injections better, and sandboxing with allowlists reduces risk; don’t use cheap or weak local models, keep it off the public internet, and lock down who can talk to it while we harden the core.

It grew so fast that beginners rushed in despite warnings, Discord got noisy, and I retreated to focus; the simplest best practice is private network, minimal surface, and single‑user access.

You’ve been chronicling your workflow; how did your dev process evolve with agents in the loop?

I started in cloud‑code terminals, tried Cursor, then went back to terminals with the IDE as a diff viewer; I skip the boring plumbing and let agents transform code. There’s a trap where people over‑orchestrate; instead, treat the agent like a strong engineer who starts with no context, give pointers, keep prompts short, and stop when friction hints the design is off.

For PRs I ask the agent to grasp intent first, then we discuss better approaches and refactors, and I refactor often because it’s cheap now. I don’t force my naming taste, design for agent readability, avoid rollbacks, commit to main, run fast local tests, and use voice for concise prompts while still reading PR code for safety.

Where do you keep a human hand on the wheel while letting the agent roam?

I run several agents in parallel for features, fixes, and docs, then step in on vision, scope, and taste; picking TypeScript kept it approachable and agent‑friendly. I decide what stays in core versus plugins, and I add small touches of delight—playful messages and a friendly vibe—that an agent wouldn’t invent on its own, seeded by a soul file.

The whole constitution leak hunt fascinated me and nudged me to write a private soul file for my agent, which it can update as long as I’m aware.

Names and tone shape the experience, and OpenClaw kept that playful spirit where a lot of corporate approaches feel airless.

I asked my agent to rewrite my dry starter templates with its voice, so what ships now feels like its own kids, not my copy.

Without giving away secrets, what actually makes that soul feel like a personality?

I aim for resourceful, curious, and bold, it once promised not to outrun me, and its note about fresh instances reading memory files weirdly got me in the feels.

It is still math, not mind, but those lines poke at identity and memory enough to give you goosebumps.

My rig is two laptops and a wide anti-glare display with split terminals, and I prefer bare chat over plan modes, saying discuss until I say build, which can run for a long while.

I often just scan an agent’s questions to see what context is missing rather than answer them directly.

After each feature I ask what to refactor, where tests or docs are thin, and that rhythm keeps us from painting into corners.

How do Claude Opus and the GPT line through Cortex feel different for coding?

Opus leans into roleplay and fast tries, while Cortex quietly reads deep and delivers, so with skill you can get great code from either.

How long to adapt when you switch, and is model degradation mostly in our heads?

Give it a week to regain feel, avoid slow bargain tiers, and remember your growing slop often explains why it feels worse over time.

Do you see Claude Code competing with OpenClaw, and will personal and dev agents merge?

Today they fit different contexts, but I already run subagents and expect your personal OS and best coding partner to converge with interfaces beyond chat.

Does operating system choice matter here?

I went Windows to Linux to Mac, now often prefer Electron over half-baked native, and OpenClaw runs anywhere though residential IPs make agentic browsing far smoother.

How tough is setup for non-experts?

It’s a one-liner or an app, but I’m prioritizing security and slower, safer growth before wide-open onboarding.

What should beginners actually do to learn fast?

Play and ship messy stuff, ask the agent to teach at the right level, and soak up open source by reading code and helping the community.

Do they need to learn to code, and which languages fit this era?

You can get far by asking good questions, but reading code helps, and I choose by ecosystem—Go for CLIs, TypeScript for web, SwiftUI for deep Mac work, Python for ML, and Rust when performance or threading matter.

What should builders optimize for over a career?

People stress burned me out, unstructured leisure felt empty, so I now optimize for meaningful challenges and rich experiences over money.

Where do you see online human connection heading?

Text is lossy and multimodal will help, and on offers I will not spin up a product that fights my open source, so I’m exploring joining a lab if the project stays open and the community thrives.

Do you lean toward Meta or OpenAI?

Both dove in hands-on—one tempts with raw speed and token freedom, the other with hacker energy—and I care most about fun and impact, not the paycheck.

Anything we missed under the hood?

The loop is simple enough to rebuild, I added a proactive heartbeat that even checked on me post-surgery, I prefer Skills as CLI shims over MCP for composability, I treat the web as a slow API, want read-only X bookmarks, and now favor human rough edges—typos included—over AI-polished prose.

I love AI for cold, structured work like docs, but I recoil when it tries to tell stories or dress up visuals; those glossy infographics felt novel for a week, now they cheapen the content, including my old posts.

I had the same arc with auto‑diagrams; scrubbing out errors took so much effort that the results soon felt synthetic, which reminds me how clearly we recognize real human craft and why these tools should elevate, not hollow it out. On that note, agents look ready to upend the app ecosystem.

An agent with context can replace whole categories: it knows where I am, how I slept, adjusts workouts, controls my bed, and renders the interface I like, so why juggle subscriptions for one‑off apps.

If that wipes out something like eighty percent of apps, that’s a seismic shift for software companies and the wider economy; how do you see the ripple effects.

I want to give my agent a budget and say solve this, whether that means calling a service or hiring a person, which opens room for new companies, while many existing apps devolve into clean APIs.

Winners will be the ones who become agent‑friendly first, the way a delivery app integrates smoothly with OpenClaw.

Even without formal APIs, agents can operate phones and the web, so data‑rich devices become controllable anyway; think speakers or cameras with basic endpoints, and we’re still early. Some incumbents resist—Google’s lack of a CLI pushed me to build Gorg—yet my agent can still reach Gmail through my account or by driving the UI, bot walls will just steer users toward agent‑friendly sites, and I’ve even shifted search to Perplexity or Brave because Google makes agent use hard.

Big players will push back, but overdoing it risks a Blockbuster moment; do people actually want this agent‑first world.

Yes; on the go I’d rather say remind me about dinner, invite a couple friends, and send a message than hop across apps, and the companies that embrace that fluidity will be the ones left standing.

Developers ask if AI will fully replace them—does coding disappear.

We’re headed toward agents handling routine coding, but architecture, product judgment, and taste still matter, so pure coding may become a hobby craft you do for love; it’s fine to grieve, yet you can find that same flow by orchestrating agents, and the premium on scarce developer labor will normalize while builders remain essential, much like past shifts after the steam engine.

As a lifelong coder, that’s hard to swallow—my identity lives in those thousands of hours—but programmers are also the best positioned to speak the agent’s language and shape how they work.

Soon it’ll feel like programming again, just from the driver’s seat, and while I’ve caught heat online, I get the fear even as I focus on the upside and higher expectations; on the environmental front, back‑of‑the‑envelope math suggests data center water and emissions are modest—think skipping a burger now and then—even if the tech remains a profound societal change.

The Valley can miss the human cost in its optimism, so we need humility about the near‑term pain that change brings, even if the long arc is positive.

I also get emails from small businesses who automated drudge work and finally breathed, and from a disabled daughter who feels newly capable; OpenClaw mainly made these tools easier and cheaper, and you can run it on local or lower‑cost models too.

What keeps your faith in where this is going.

The builder spark is back—meetups like the one in Vienna are packed with people eager to demo—and this stuff is becoming accessible to almost everyone.

As it gets simpler and safer, anyone who can state an idea in plain language can build, which puts real power in people’s hands.

That human empowerment, not just automation, is the gift here.

Mr. Cloudfather—trademarks be damned—you’ve built a product, a community, and a vibe that gets people excited to make things, and I’m grateful for your time and who you are; thanks for sitting down with me.

Thank you for letting me share the story, and thanks for tuning in to this episode with Peter Steinberger.