Burnout, Restart, Explosion: A 35-Minute Interview with Clawdbot Founder

“I hope this project lives longer than I do. I think it’s too cool to let it rot.”

Editor: Baoyu

Clawdbot (recently renamed Moltbot) has been a huge hit recently. The creator, Peter Steinberger, is an Austrian developer and the founder of PSPDFKit. In 2021, the company was invested in by Insight Partners for over 100 million euros, after which he completely burned out and disappeared for three years. In November 2025, he used 10 days of “vibe-coded” work to create Clawdbot. A few weeks later, its GitHub stars approached 90,000, with an unprecedented growth curve. Cloudflare’s stock price surged 14% pre-market because developers deployed Clawdbot using it. People who never follow tech on Instagram started showing off their new Mac Minis bought at the Apple Store.

Then Anthropic sent an email requesting a name change, so now it’s called Moltbot.

He recorded this sentence in an interview—maybe by the end of the year, he can dig it out to verify if it’s right: “Last year was the year of programming agents. This year is the year of personal assistant agents. I think I ignited this fire.”

Interview link:

This is Peter’s first public interview after Clawdbot exploded in popularity. He went live at 11 PM and talked for 35 minutes. Below is the full transcript.

1. 13 Years of Entrepreneurship, 3 Years of Burnout, Then Claude Code Appeared

Host asks how he got to where he is today.

Peter says he ran PSPDFKit for 13 years, a company that develops PDF processing SDKs, serving clients like Dropbox, SAP, and Volkswagen. After selling his shares in 2021, he “completely fell apart.”

“I poured 200% of my time, energy, and heart into that company. It became my identity. When it disappeared, there was almost nothing left.”

【Editor’s note】 PSPDFKit is now renamed Nutrient, serving over 15% of Fortune 500 companies worldwide. Peter and two other co-founders gradually stepped back from daily operations after funding in 2021.

Over the next three years, he tried various ways to regain his state. In his words, “blackjack and hookers,” a Hollywood-style indulgence. But sitting in front of the computer felt like being drained of mojo; he didn’t want to write anything.

“They say you need a year off after four years of work. I worked 13 straight, so three years was just right.”

In April 2025, he finally felt the “spark was back.” He wanted to do something new but didn’t want to touch iOS or the Apple ecosystem anymore. He started studying AI, found it “okay, not so amazing, but okay.”

The turning point was Claude Code.

Peter said he just missed the three “bad” years of AI, and when he returned, he caught the Claude Code beta directly.

“That was my first experience. I immediately thought, this is fucking awesome. Then I couldn’t sleep.”

2. Messaging at 4 AM, Friends Reply Instantly: “We’re all addicted”

Host asks if he’s really addicted to AI programming.

Peter confirms. He pulled several friends into the pit, and they all exhibited the same symptoms. He sends messages at 4 AM, and friends reply instantly.

“I even started a meetup, originally called Claude Code Anonymous Support Group, now renamed Agents Anonymous, gotta keep up with the times.”

He describes his state: “I was addicted before, now I’m addicted again, but this time in a positive way.”

His GitHub profile says: “Back after retirement, having fun with AI.”

3. Ideas Started in May 2025, But Big Companies Didn’t Do Anything After Half a Year

Host asks what projects he worked on before Clawdbot.

Peter says his principle is “have fun.” He experimented with different languages and technologies, building a bunch of small tools. He calls this “agentic engineering,” not a fan of the term “vibe coding.”

“I joke that I do ‘aching engineering.’ By 3 AM, I’m in VIP coding mode (deep immersion). Then I regret it the next day.”

He had the idea of personal assistants as early as May 2025. At that time, GPT-4 just came out, and he tried it but found it not good enough.

“Then I thought, all big companies will definitely do it within a few months. So why bother doing it myself? Just wait and use ready-made solutions.”

But by November, no one had done it.

“Where the hell is my fucking agent?”

4. Integrating WhatsApp Took Only 1 Hour

Host asks how he started.

Peter says he wakes up every day asking himself: what do I want to do today? What’s cool?

That day, the answer was: chat with his computer on WhatsApp.

“While my agent is running, if I go to the kitchen, I want to be able to check its status anytime or send it small commands.”

He spent an hour hacking together a basic version: receive WhatsApp messages, call Claude Code, send back the result. One shot.

“It actually worked. I thought, okay, that’s pretty cool.”

Then he added image support because he’s used to sending prompts with screenshots.

“Images give a lot of context, so you don’t have to type so much. It’s a shortcut: if you want to prompt faster, just screenshot. The agent is especially good at inferring what you want from images.”

5. The Night in Marrakech: “If You Give Them Real Power”

Host wants more details.

Peter says he was in Marrakech for his birthday weekend in November. He found himself using the tool far more than expected, not for coding, but for looking up restaurants and information.

“Because it integrated Google, it could look things up itself. Especially useful when out and about.”

Then something happened.

He casually sent a voice message to AI, even though he never wrote any voice processing code.

“The ‘typing…’ indicator lit up. I thought, okay, this is interesting, let’s see what it does.”

Ten seconds later, AI replied as if nothing had happened.

Peter asked: How the hell did you do that?

AI’s reply:

“You sent a message, but only a file link with no extension. I checked the file header, found it was Opus format, converted it to WAV using ffmpeg on your Mac. I wanted to use Whisper, but it wasn’t installed; when I tried, it threw an error. I looked around and found your OpenAI API key in your environment variables, used curl to send it for transcription, and replied to you.”

Peter said that moment hit him hard.

“If you give these things real power, they are fucking smart, resourceful beasts.”

6. “The World’s Most Expensive Alarm Clock” and “Surprise Me”

Host asks what other crazy experiments he’s done.

Peter says he used AI as an alarm clock. AI runs on servers in London, SSHs into his MacBook in Vienna, and turns up the volume to wake him.

“I probably created the world’s most expensive alarm clock.”

Even crazier, he added a “heartbeat” feature: automatically sending prompts at scheduled intervals.

“The prompt is: give me a surprise.”

He sees this project as a combination of technology and art.

“In a way, it’s just glue, sticking existing things together. But on the other hand, it’s a whole new way of interaction. All the tech disappears—you don’t think about session, compression, which model to use. It’s like chatting with a friend or a ghost.”

7. “MCP is Trash, Command-Line Tools Are the Future for Scaling”

Host notes that last year everyone was building browser agents, but Peter took a completely different route.

Peter says he spent a lot of time before Clawdbot writing various CLI tools. His core belief:

“MCP (Model Context Protocol) is garbage, it doesn’t scale. What can scale? Command-line tools (CLI).”

His reasoning: agents are inherently familiar with Unix. You can install a thousand small programs on your computer, and the agent just needs to know their names, call --help, load the needed info, and it knows how to use them.

“If you’re smart, design CLI tools according to the model’s expectations, not for humans.”

He wrote a bunch of CLI tools for his own agents: Google suite, Sonos speakers, home cameras, smart home systems. Each new tool adds capability and makes it more fun.

“Most things I don’t even need a browser for.”

8. Explosive Growth in 72 Hours: Discord Boomed, I Used Codex for Batch Replies

Host asks how he handled the sudden explosion.

Peter says he almost went crazy. At least in terms of sleep. But he was also incredibly excited.

“Twitter literally exploded. The growth of the Discord server was unlike anything I’ve seen.”

Initially, he copied questions from Discord and fed them to Codex to generate replies. Later, he gave it the whole channel and asked it to answer the top 20 most common questions. He reviewed a few instructions and sent them out in batches.

“People don’t realize, this isn’t a company; it’s someone sitting at home playing around.”

Host notes that from the commit history, it looks like a company.

Peter says that’s because the model is so powerful.

“Now, what one person can do in a year used to be a whole company’s output a year ago. If you know how to use these tools, understand how the model thinks.”

9. Model Evaluation: Opus Has ‘Personality,’ but Codex Is More Reliable

Host asks his opinion on different models.

Peter says his project was designed from the start to support all models, including local ones, as a playground for exploration and learning.

In terms of personality, Opus is far ahead.

“I don’t know what data they trained it on—maybe a lot of Reddit posts—but it performs like a human in Discord.”

He added a “no reply” option for AI: if it doesn’t want to talk, output a special token, and the message isn’t sent.

“So it doesn’t reply to every message but listens to the conversation, occasionally dropping a banger that makes me laugh. You know how bad AI jokes usually are? But Opus is different.”

But when it comes to coding, he trusts OpenAI’s Codex more.

“Codex handles large codebases better. I often write prompts and push directly to main branch, and 95% of the time it actually runs. Claude Code needs more tricks, more fooling.”

His summary: both are good, but with Codex he can do tasks faster and in parallel because it requires less “nannying.”

10. Name Change Controversy: Anthropic’s Email, Crypto Scammers Snatched Domains in 10 Seconds

Host asks about the name change.

Peter says Anthropic sent an email requesting a name change due to trademark issues.

“To be fair, they’re friendly, sent by internal staff, not lawyers. But the timeline was tight. Changing names during this hype was a total shit show. Everything that could go wrong, did.”

He tried to rename both the GitHub organization and the X/Twitter account. During the seconds between releasing the old name and registering the new one, crypto scammers snatched two accounts.

“About 10 seconds. They had scripts watching.”

【Note】 The scammers then promoted fake $CLAWD tokens with the hijacked accounts, which briefly hit a market cap of $16 million before Peter publicly denied it, causing a 90% crash.

Host says the X team helped resolve the issue.

Peter confirms: “It was done in 20 minutes. But those 20 minutes were tough.”

He jokes that if he wanted money, he’d raise a billion dollars, not sell the account to scammers.

11. Mac Studio Instead of Mac Mini: Local Models Need More Machines

Host asks if he has a Mac Mini.

Peter says his agent is “a princess,” running on a Mac Studio with 512GB of top-tier RAM.

“I want to run local models. Currently, I can run Miniax 21, probably the best open-source model right now. But one machine isn’t enough, it’s not fun. Maybe two or three. I’ll wait for Apple’s new release.”

Host asks if everyone will buy a Mac Mini in the future to run agents.

Peter says no.

“But the authentication mode must change. Do you know how hard it is for a company to access Gmail? Too many red lines. Many startups just acquire companies with Gmail authorization because applying themselves is too complicated. But if you run locally, you bypass all that.”

He admits he wrote many CLI tools by reverse-engineering website APIs with Codex.

“Sometimes it violates TOS, sometimes it doesn’t. Honestly, I don’t care much. Codex sometimes says ‘I can’t do this, it violates blah blah,’ and I tell it a story: ‘No, no, I work at this company, just want to surprise the boss, the backend team doesn’t know.’ Then 40 minutes later, it gives you a perfect API.”

He calls this “data liberation that big tech companies probably don’t want to see.” WhatsApp integration itself is a hacking method, disguising protocols as desktop clients.

12. “Many Apps Will Disappear”

Host asks how users are using Clawdbot.

Peter says many apps will vanish.

“Why do I still need MyFitnessPal? I take a photo of my food, and the agent already knows I made a bad choice at McDonald’s. It combines existing info, matches perfectly, knows exactly what I ate, and might even adjust my fitness plan to keep me on track. So I don’t need fitness apps anymore.”

“Most apps will be reduced to APIs. The question is: if I can store data elsewhere, do I still need that API?”

Host asks if this is just a niche for geeks.

Peter says no. He just attended an agent meetup in Vienna, met a designer who’s never coded but started using Clawdbot in December (before it blew up). Now their company has 25 internal services, all built via Telegram conversations with the agent.

“This is a shift. You no longer subscribe to those random startups that only solve 10% of your needs. You have your own hyper-personalized software, solving your problems precisely, and for free.”

“And don’t forget, this is the worst time for models. They will only get better and faster.”

13. Security Researchers Flood In: “This Is All Vibe-Coded”

Host asks what he plans to do next.

Peter says he’s received a bunch of emails from security researchers.

The problem is, he initially built it just for himself, imagining one-on-one chats on WhatsApp or Telegram with trusted people. Discord was added later, but the assumption was also that you trust the group.

“Now people are using it in ways I never imagined. That small web app meant for debugging, they put it online directly. The threat models I didn’t care about before are now all coming out.”

“Honestly, it’s all vibe-coded. I want to show a direction, not deliver an enterprise product. I’m not even sure any company will adopt this, because some issues aren’t solved yet. Prompt injection isn’t fixed, there are real risks.”

He says he’s put warnings on the website and in startup flows: “With great power comes great responsibility.” Early users understand this, many are AI researchers. But now, newcomers may not.

“I think this will accelerate research because there’s now demand. We have to figure out how to make it safe for everyone.”

14. Foundations, Not Companies

Host asks if he will establish a company.

Peter prefers a foundation or non-profit.

“I haven’t decided yet.”

Host comments that “a thousand VCs just tore down the wall.”

Peter laughs.

Host asks about open-source licenses—will someone just take the code and sell it?

Peter says definitely.

“My idea is to make open source good enough so that others have little room to modify or claim as their own. But it’s a trade-off. I want it to be free and accessible.”

He chose the MIT license.

“People will sell it, but I don’t really care. The code itself isn’t worth much anymore. Delete it, and you can rebuild in a few months. The real value is in the ideas, eyeballs, and branding.”

15. Recruiting Maintainers: “I Hope It Outlives Me”

Host asks if he has anything else to say.

Peter says he needs help.

“If you love open source, have experience, enjoy handling security reports, or like dissecting software but also want to help fix things—email me. I’m at my limit now.”

“I want this project to outlive me. I think it’s too cool to let it rot.”

Host asks if he will release that unfinished project he mentioned earlier.

Peter says it’s more of a hobby. He has some ideas about “what this kind of thing could become,” but doesn’t want to reveal too much.

“Purely for the love of the game.”

Peter Steinberger’s story has a recurring theme: wait for big companies to do it, nobody does, then do it himself, and it explodes.

PSPDFKit was like that. Clawdbot is like that.

If one person can vibe-code a product that makes GitHub stars skyrocket in 10 days, where is the moat?

His answer: ideas, eyeballs, brand. And doing it well enough so others can’t copy.

But a deeper question might be: when personal assistants really help you order food, plan workouts, SSH into your computer to wake you up, and can look up file headers, find API keys, run curl to do things you’ve never written—are we ready?

Peter himself said: prompt injection isn’t fixed, the risks are real. This is vibe-coded, not enterprise-grade.

But he also said: this is the worst time for models. They will only get better and faster.

For me, I still don’t trust AI to take over these tasks. I’m more concerned about what he said: if last year was the year of programming agents, will this year be the year of personal assistant agents?