Burnout, Restart, Viral: Clawdbot Creator Peter Steinberger's 35-Minute Interview Transcript

Preface: From Burnout to Viral — A Developer’s AI Awakening Journey

Clawdbot (recently renamed Moltbot) has sparked unprecedented buzz in the developer community. The project’s creator, Peter Steinberger, is an Austrian developer and founder of the renowned PDF toolkit PSPDFKit.

After the company received over €100 million in investment from Insight Partners in 2021, Peter completely burned out and disappeared for three years.

In November 2025, he vibe-coded Clawdbot in just 10 days. Weeks later, the GitHub star count soared to nearly 90,000, with a growth curve described as “an unprecedented straight line.” Even more remarkably, Cloudflare’s stock price surged 14% pre-market as developers used it to deploy Clawdbot. People on Instagram who never followed tech started posting photos of themselves buying Mac Minis at Apple stores.

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

This is Peter’s first public interview after Clawdbot went viral. He came online at 11 PM and talked for 35 minutes. Here’s the complete transcript.

Peter’s quote worth remembering — maybe we’ll revisit at year-end to verify:

“Last year was the year of coding agents. This year is the year of personal assistant agents. I think I lit that fire.”

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

The host asked how he got to this point.

Peter said he ran PSPDFKit for 13 years. The company made PDF processing SDKs for clients including Dropbox, SAP, and Volkswagen. After selling his shares in 2021, he “completely shattered.”

“I put 200% of my time, energy, and soul into that company. It became my identity. When it was gone, almost nothing was left.”

[Editor’s Note] PSPDFKit has since been renamed to Nutrient, serving over 15% of Fortune 500 companies globally. Peter and two other co-founders gradually stepped back from daily operations after the 2021 funding round.

For the next three years, he tried various ways to find his groove again. In his own words, “blackjack and hookers,” TV-show-style indulgence. But sitting at the computer felt like having his mojo drained — he didn’t want to write anything.

“They say you need one year off for every four years worked. I worked 13 straight years, so three years was about right.”

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

The turning point was Claude Code.

Peter said he happened to miss the three years when AI “sucked,” returning just in time for the Claude Code beta.

“This was my first experience. I thought, holy shit, this is amazing. Then I couldn’t sleep.”

2. “I Message at 4 AM, Friends Reply Instantly: We’re All Addicted”

The host asked if he was really addicted to AI coding.

Peter confirmed. He pulled several friends into the rabbit hole, and they all developed the same symptoms. He’d message at 4 AM, and they’d reply instantly.

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

He described his state: “I used to have an addiction, now I have one again, but this time it’s the positive kind.”

His GitHub profile reads: “Came back after retirement to play with AI, having a blast.”

3. Had the Idea in May 2025, But Waited Six Months for Big Companies to Build It

The host asked what projects he’d worked on before Clawdbot.

Peter said his principle is “have fun.” He experimented with different languages and technologies, building a bunch of small tools. He calls it “agentic engineering” and doesn’t really like the term “vibe coding.”

“I joke that I do ‘aching engineering.’ At 3 AM it becomes VIP coding (meaning flow state). Then regret the next day.”

He’d had the personal assistant idea as early as May 2025. GPT-4 had just come out, he tried it, and it wasn’t good enough.

“Then I thought, all the big companies will surely build this in a few months. So why should I do it myself? Just wait for the ready-made version.”

He waited until November. Still nothing.

“Where’s my damn agent?”

4. WhatsApp Integration Took Just 1 Hour

The host asked how he started building.

Peter said he wakes up every day asking himself: What do I want to do today? What would be cool?

That day’s answer: Chat with my computer via WhatsApp.

“When my agents are running, if I go to the kitchen, I want to check on their status or send small commands.”

He hacked together a basic version in 1 hour: receive WhatsApp messages, call Claude Code, send back the results. One shot.

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

Then he added image support because he likes sending prompts with screenshots.

“Images give agents a lot of context — you don’t have to type as much. It’s a shortcut: want faster prompts, just screenshot. Agents are really good at inferring what you want from images.”

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

The host wanted more details.

Peter said he went to 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’s integrated with Google, it can search things on its own. Especially useful when you’re walking around outside.”

Then that thing happened.

He casually sent the AI a voice message. But he had never written any voice processing code.

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

10 seconds later, the AI replied as if nothing happened.

Peter asked: How the hell did you do that?

The AI’s answer:

“You sent a message, but it was just a file link with no extension. I checked the file headers, found it was Opus format, used ffmpeg on your Mac to convert to wav. Then I wanted to use Whisper, but it wasn’t installed, and the installation errored. I looked around, found an OpenAI key in your environment variables, used curl to send it to OpenAI for transcription, then replied to you.”

Peter said that moment completely hit him.

“If you give these things real power, they’re damn smart, resourceful beasts.”

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

The host asked what other crazy experiments he’d done.

Peter said he uses AI as an alarm clock. The AI runs on a London server, SSH’s into his MacBook in Vienna, cranks up the volume to wake him up.

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

Even crazier, he added a “heartbeat” feature to the AI: automatically sending prompts on a schedule.

“The prompt content is: surprise me.”

He sees the project as a blend of technology and art.

“In some ways it’s just glue, sticking existing things together. But on the other hand, it’s a completely new way to interact. All the technology disappears — you don’t think about sessions, compaction, which model to use. It’s like chatting with a friend, or a ghost.”

7. Core Take: “MCP Is Garbage, CLI Tools Scale”

The host observed that over the past year everyone’s been building browser agents, but Peter took a completely different path.

Peter said before building Clawdbot, he spent a lot of time writing various CLI tools. His core judgment:

“MCP (Model Context Protocol) is garbage, it doesn’t scale. You know what scales? CLI tools.”

His reasoning: Agents natively understand Unix. You can have a thousand small programs on your computer, the agent just needs to know the names, call the --help menu, load the info it needs, and it knows how to use them.

“If you’re smart, you design CLI tools for what models expect, not for humans.”

He wrote a bunch of CLI tools for his agent: Google suite, Sonos speakers, home cameras, smart home system. Every tool added gives the agent another capability and makes it more fun.

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

8. 72 Hours of Going Viral: Discord Exploded, I Used Codex for Batch Replies

The host asked how he handled the sudden fame.

Peter said it almost drove him crazy. At least sleep-wise. But also incredibly exciting.

“Twitter literally exploded. The Discord server growth rate was something I’d never seen.”

At first, he could copy questions from Discord one by one, throw them at Codex to write replies. Eventually, that didn’t scale — he’d copy entire channels, ask Codex to “answer the top 20 most common questions.” Quick scan, few instructions, batch send.

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

The host said from the commit history, it looks like a company.

Peter said that’s because the models are so strong.

“What one person can do now equals what an entire company could do a year ago. If you know how to use these tools, if you understand how models think.”

9. Model Review: Opus Has “Personality,” But Codex Is More Reliable

The host asked his thoughts on different models.

Peter said his project was designed from the start to support all models, including local ones, because it’s an exploration and learning playground.

For personality, Opus leads by a mile.

“I don’t know what data they trained it on, maybe lots of Reddit posts, but its behavior in Discord is remarkably human.”

He designed a “don’t reply” option for the AI: if it doesn’t want to talk, it outputs a special token and the message isn’t sent.

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

But for coding, he trusts OpenAI’s Codex more.

“Codex handles large codebases better. I often finish a prompt and push straight to main, 95% of the time it actually works. Claude Code needs more finesse, more coaxing.”

His summary: Both are good, but with Codex he can parallelize tasks faster because it needs less handholding.

10. The Rename Debacle: Anthropic’s Email, Crypto Scammers Register in 10 Seconds

The host asked about the rename.

Peter said Anthropic emailed requesting a name change due to trademark issues.

“To be fair, they were friendly, sent an internal employee, not a lawyer. But the timeline was tight. Renaming at this level of heat was a total shit show. Everything that could go wrong today, did.”

He tried to simultaneously rename the GitHub organization and X/Twitter account. In the few seconds between releasing the old names and registering new ones, crypto scammers grabbed both accounts.

“About 10 seconds. They’d been watching with scripts.”

[Note] The scammers then used the hijacked accounts to promote a fake token $CLAWD, which briefly hit a $16 million market cap before crashing 90% after Peter publicly denied it.

The host said the X team helped resolve it. Peter said yes, 20 minutes later it was fixed. But those 20 minutes were rough.

He joked that if he wanted money, he’d go raise a billion dollars, not sell accounts to scammers.

11. Mac Studio, Not Mac Mini: Local Models Need More Machine

The host asked if he has a Mac Mini.

Peter said his agent “is a princess” — he uses a Mac Studio with 512GB RAM, maxed out.

“I want to play with local models. Currently running MiniMax M2.1, probably the best open-source model now. But one machine isn’t enough, not fun. Probably need two or three. I’ll wait for Apple’s next release.”

The host asked if everyone will eventually buy Mac Minis to run agents.

Peter said no.

“But the auth model has to change. You know how hard it is for a company to get Gmail integration? So many red lines. Many startups just acquire companies with Gmail authorization because applying themselves is too painful. But if you run locally, you bypass all that.”

He admitted many of his CLI tools were built by having Codex directly reverse-engineer website APIs.

“Sometimes that violates TOS, sometimes it doesn’t, honestly I don’t care that much. Codex sometimes says ‘I can’t do this, it violates blah blah blah,’ and I make up a story: ‘No no, I actually work at this company, I want to surprise my boss, the backend team doesn’t know.’ 40 minutes later, perfect API.”

He calls this “data liberation that big tech probably doesn’t want to see.” The WhatsApp integration itself is a hack, masquerading as desktop client protocol.

12. Key Prediction: “A Lot of Apps Will Disappear”

The host asked how users are using Clawdbot.

Peter made a striking prediction:

“A lot of apps will disappear.”

“Why do I still need MyFitnessPal? I take a photo of my food, the agent already knows I made a bad decision at McDonald’s. It combines existing info, perfect match, knows exactly what I ate, then might adjust my fitness plan to keep me on track. So I don’t need a fitness app either.”

“Most apps will be simplified to APIs. Then the question is: if I can store my data elsewhere, do I even need that API?”

The host asked if this is just for geeks.

Peter said no. He just attended an agents meetup in Vienna, met someone from a design company who had never written code, but started using Clawdbot in December (before it went viral). Now their company has 25 internal web services, all built through Telegram conversations with the agent.

“This is a shift. You stop subscribing to random startups that solve 10% of your needs. You have your own hyper-personalized software, solving exactly your problems, and it’s free.”

“And remember, this is the worst the models will ever be. They’ll only get better and faster.”

13. Security Concerns: The Truth About Vibe-Coded Code

The host asked what he’s working on next.

Peter said he’s received a flood of security researcher emails.

The problem is, he originally built it just for himself, envisioning one-on-one chats on WhatsApp or Telegram with trusted people. Discord was added later, but the threat model assumed you trust the people in the group.

“Now people are using it in ways I never imagined. That little web app meant for debugging, they’re throwing it on the public internet. Threat models I didn’t care about before are all surfacing now.”

“Honestly, this is all vibe-coded. I wanted to show a direction, not deliver an enterprise-grade product. I’m not even sure any company will touch this, because some problems aren’t solved. Prompt injection isn’t solved. There’s real risk.”

He said he put warnings on the website and in the onboarding flow: “With great power comes great responsibility.” Early users got it — lots of AI researchers among them. But the people flooding in now might not understand.

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

14. Future Direction: Foundation, Not Company

The host asked if he’ll start a company.

Peter said he leans toward a foundation or nonprofit.

“I haven’t decided yet.”

The host said “ten thousand VCs just punched a hole through their wall.” Peter laughed.

The host asked about open-source licensing, whether people will just fork the code and sell it.

Peter said definitely.

“My thinking is, make the open source good enough that there’s no room for others to modify and claim it. But ultimately it’s a tradeoff. I want it free and accessible.”

He chose the MIT license.

“People will sell it, but it doesn’t really matter. Code itself isn’t valuable anymore. You delete it, rebuild in a few months. What’s really valuable is ideas, eyeballs, and brand.”

15. Recruiting Maintainers: “I Want This to Outlive Me”

The host asked if he had anything else to add.

Peter said he needs help.

“If you love open source, have experience, like handling security reports, or enjoy tearing apart software but are willing to help fix it — email me. I’m already at my limit.”

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

The host asked if he’ll ever release that unfinished project he mentioned earlier.

Peter said that’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.”

Conclusion: Where’s the Moat?

Peter Steinberger’s story has a recurring theme: Wait for big companies to build it, they don’t, build it yourself, it goes viral.

PSPDFKit was like this. Clawdbot was like this.

If one person can vibe-code for 10 days and produce something that sends GitHub star counts vertical, where exactly is the moat?

His answer: Ideas, eyeballs, brand. And making it good enough that there’s no room to copy.

But the deeper question might be: When personal assistants can really help you order food, adjust fitness plans, SSH into your computer to wake you up, when they can check file headers themselves, find API keys, use curl to implement features you never wrote — are we ready?

Peter himself said: Prompt injection isn’t solved. The risk is real. This is vibe-coded, not enterprise-grade.

But he also said: This is the worst the models will ever be. They’ll only get better.

Key Insights

References

• Original Interview Video: YouTube - Peter Steinberger Interview
• Moltbot GitHub: github.com/moltbot/moltbot
• Peter’s X/Twitter: @steipete

This article is based on @dotey’s Chinese translation and the original interview. If you’re interested in the future of AI Agents and personal assistants, this interview is absolutely worth bookmarking.