Clawdbot Founder Peter Steinberger on Burnout, AI Addiction, and the Future of Personal Agent Tools

Organized by: Baoyu

Clawdbot (recently renamed Moltbot) has recently gone viral. Its creator, Peter Steinberger, is an Austrian developer and the founder of PSPDFKit. In 2021, the company received over 100 million euros in investment from Insight Partners, after which he completely burned out and disappeared for three years. In November 2025, he built Clawdbot in just 10 days "by vibe" (vibe-coded). Within weeks, the project had nearly 90,000 stars on GitHub, with a growth curve described as "a straight line never seen before." Cloudflare's stock price surged 14% pre-market as developers deployed Clawdbot using their services. On Instagram, even people who never paid attention to technology started posting pictures of themselves buying Mac Minis from Apple Stores.

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

He said this in the interview, so I'll note it down first. Maybe by the end of the year, we can bring it up again to see if it's correct: "Last year was the year of programming agents (Agents), and this year is the year of personal assistant agents. I think I lit this fire."

Interview Address:

The link you provided is a YouTube video URL. If you'd like, I can help you find the title or description of the video, or assist with anything else related. Let me know what you need!

This is Peter's first public interview after the Clawdbot project went viral. He went online at 11 PM and talked for 35 minutes. Below is the full transcript.

The host asked him how he had gotten to this point today.

Peter said he had been running PSPDFKit for 13 years. The company provides PDF processing development toolkits (SDKs), with clients including Dropbox, SAP, and Volkswagen. After selling his shares in 2021, he "completely fell apart."

"I put 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 has now been renamed to Nutrient, serving over 15% of Fortune 500 companies globally. After the 2021 funding round, Peter and the two other co-founders gradually stepped back from day-to-day operations.

For the next three years, he tried every possible way to regain his form. As he put it himself, "eating, drinking, gambling (blackjack), and hookers"—a TV-show style of indulgence. But sitting in front of the computer felt like having all his vitality (mojo) drained away; he didn't feel like writing anything at all.

"They say you need to take a year off for every four years of work. I worked continuously for 13 years, so three years is just right."

In April 2025, he finally felt that the "spark had returned." He wanted to try something new, but didn't want to work with iOS or the Apple ecosystem anymore. He started exploring AI and found it "okay—not particularly impressive, but still okay."

The turning point was Claude Code.

Peter said he just missed the three years when AI was "terrible," and when he came back, he immediately joined the Claude Code beta.

This was my first experience. At that moment, I just thought, "This is fucking awesome!" Then I couldn't fall asleep.

The host asked him if he was really addicted to AI programming.

Peter said yes. He pulled several friends into the pit, and as a result, everyone developed the same symptoms. He sent a message at 4 a.m., and his friends replied instantly.

I even started a gathering, originally called the "Claude Code Anonymous Support Group," and it's now renamed to "Agents Anonymous" to keep up with the times.

He described his current state: "I used to have an addiction, and now I have one again, but this time it's a positive kind."

His GitHub homepage says: "Back to play with AI after retirement, having a lot of fun."

The host asked what projects he had worked on before Clawdbot.

Peter says his principle is "have fun." He experiments with different languages and technologies, building a bunch of small tools. He calls this "agentic engineering," and he's not a big fan of the term "vibe coding."

I jokingly said that what I do is "aching engineering." By 3 a.m., I'm in VIP coding mode (meaning a state of deep immersion). Then I regret it the next day.

The idea of a personal assistant had already occurred to him back in May 2025. At that time, GPT-4 had just been released, and he tried it out, but found it wasn't good enough yet.

Then I thought, all the big companies will definitely come up with it within a few months. So why should I make it myself? It's better to just wait and use the ready-made one.

But by November, no one had done it yet.

"Where the hell is my agent?"

The host asked him how he got started.

Peter says that every time he wakes up, he asks himself: What do I want to do today? What would be cool?

The answer that day was: Chatting with your computer on WhatsApp.

"When my agents are running, if I go to the kitchen, I want to be able to check their status anytime or send them some small commands."

He spent 1 hour hacking together a basic version: receiving WhatsApp messages, calling Claude Code, and sending back the results. Done in one go (One shot).

"It actually worked right away. I thought, well, that's pretty cool."

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

Images can provide a lot of context to an AI agent, so you don't need to type as much. Here's a shortcut: if you want to generate prompts more quickly, just take a screenshot. AI agents are especially good at inferring what you want from images.

The host wants to hear more details.

Peter said he went to Marrakech for a birthday weekend in November. He found himself using this tool much more frequently than he had expected, and not for writing code, but for looking up restaurants and information.

"Because it's integrated with Google, it can look things up by itself. It's especially useful when you're out and about."

Then that thing happened.

He casually sent a voice message to the AI. But he hadn't written any voice processing code at all.

The "typing" indicator lit up. I thought, great, now it gets interesting—let's see what it does.

After 10 seconds, the AI responded as if nothing had happened.

Peter asked it, "How the hell did you do it?"

AI's response:

"You sent a message, but it only contained a file link without an extension. I checked the file header and found it to be in Opus format, so I used ffmpeg on your Mac to convert it to wav. Then I wanted to use Whisper, but it wasn't installed and there was an error during installation. I looked around and found OpenAI's API key in your environment variables, so I used curl to send it to OpenAI for transcription and then replied to you."

Peter said that at that moment, he was completely struck.

"If you give these things real power, they're damn clever, resourceful beasts."

The host asked him about other crazy experiments he had done.

Peter said he uses AI as an alarm clock. The AI runs on servers in London, logs into his MacBook in Vienna via SSH, and wakes him up by increasing the volume.

"I probably made the most expensive alarm clock in the world."

Even more crazily, he added a "heartbeat" feature to the AI: automatically sending prompts at regular intervals.

The prompt is: Surprise me.

He regards this project as a combination 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 of interacting. All the technology disappears—you don't have to think about session state, compaction, or which model to use. It's as if you're chatting with a friend, or a ghost.

The host noted that everyone has been working on browser agents over the past year, but Peter took a completely different approach.

Peter said that before working on Clawdbot, he spent a lot of time writing various command-line tools (CLI). His core judgment is:

"MCP (Model Context Protocol) is garbage and can't scale. Do you know what can scale? Command-line tools (CLI)."

His reasoning: agents are naturally familiar with Unix. You can install a thousand small programs on your computer, but an agent only needs to know their names, invoke the --help menu, load the necessary information, and then it knows how to use them.

"If you're smart, you'll design the command-line tool according to the model's expectations, rather than for humans."

He wrote a bunch of command-line tools for his agent: the Google suite, Sonos speakers, home cameras, and the smart home system. Every time he added a new tool, the agent gained another capability and became more fun.

"I don't really need a browser for most things."

The host asked him how he coped with sudden fame.

Peter said he's almost going crazy. At least when it comes to sleep. But he's also incredibly excited.

"Twitter literally exploded. The growth of Discord servers is faster than I've ever seen."

At first, he could still copy questions from Discord one by one and feed them to Codex to generate responses. Later, he couldn't even do that anymore. Instead, he directly copied the entire channel and told Codex, "Answer the 20 most common questions." He would skim through them, give a few instructions, and then send them out in bulk.

"People don't realize that this isn't a company, but just a person playing by themselves at home."

The host said that, looking at the commit history, it seems very much like a company.

Peter said that it's because the model is too strong.

What one person can accomplish now is equivalent to the output of an entire company a year ago. If you know how to use these tools and understand how the models think.

The host asked for his opinions on different models.

Peter said that his project was initially designed to support all models, including local models, because it serves as an experimental playground for exploration and learning.

In terms of personality, Opus is far ahead.

"I don't know what data they trained it on, whether there are a lot of Reddit posts, but it behaves so human-like in Discord."

He designed a "do not reply" option for AI: if it does not want to respond, it outputs a special token, and the message is not sent.

"So it doesn't respond to every message, but instead listens to the conversation and occasionally drops a brilliant reply (a 'banger') that makes me laugh. You know how bad AI jokes usually are. But Opus is different."

But when it comes to writing code, he trusts OpenAI's Codex more.

"CodeX has a stronger capability in handling large codebases. I often write the prompt and directly push it to the main branch, and it actually works in 95% of cases. Claude Code requires more skill and more coaxing."

His summary: Both are good, but Codex allows for faster parallel tasks because it requires less "hand-holding."

The host asked him about the name change.

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

To be fair, they were very friendly and sent internal staff rather than lawyers. But the timeline was extremely tight, and changing the name amid such high public attention was indeed a complete mess (a shit show). Everything that could go wrong today did go wrong.

He tried to rename both his GitHub organization and his X/Twitter account at the same time. In the few seconds between releasing the old names and registering the new ones, cryptocurrency scammers snatched up both accounts.

About 10 seconds. They've been monitoring with scripts long ago.

[Note] The scammer then used a newly registered account to promote the fake token $CLAWD, which briefly reached a market cap of $16 million. After Peter publicly denied any association, the value of the token plummeted by 90%.

The host said that Team X helped him resolve the issue.

Peter said yes, and it would be done in 20 minutes. But those 20 minutes were hard to wait through.

He also joked that if he wanted money, he would raise $1 billion, not sell his account to scammers.

The host asked him if he had a Mac Mini.

Peter said his agent "is a princess," and he's using a Mac Studio with the top configuration of 512GB of memory.

"I want to play with a local model. I can currently run Miniax 21, which is probably the best open-source model available now. But one machine isn't enough—it's not very fun. I might need two or three machines. I want to wait for Apple's new model before deciding."

The host asked, "In the future, will everyone buy a Mac Mini to run AI agents?"

Peter said no.

"But the authentication (auth) model must change. Do you know how difficult it is for a company to integrate with Gmail? There are so many restrictions that many startups directly acquire companies that already have Gmail authorization, because applying for it themselves is too complicated. But if you run it locally, you bypass all of these issues."

He admitted that he had written many command-line tools by directly reverse-engineering website APIs using Codex.

"Sometimes it violates the terms of service (TOS), sometimes it doesn't, and to be honest, I don't really care. Codex will sometimes say, 'I can't do this, it violates blah blah blah,' so I'll make up a story for it: 'No no, I actually work at this company, and I want to surprise my boss, the backend team doesn't know.' Then after 40 minutes, it gives you a perfect API."

He said this is "data liberation that big tech companies might not want to see." The WhatsApp integration itself is also a hack, mimicking the protocol of a desktop client.

The host asked him about what he observed regarding how users used Clawdbot.

Peter said that many apps will disappear.

"Why do I even need MyFitnessPal? I just take a photo of my food, and the AI already knows I made a wrong choice at McDonald's. It combines the available information, perfectly matches it, and precisely knows what I've eaten. It might even adjust my fitness plan to help me stay on track. So I don't need a fitness app either."

Most apps will be simplified into APIs. Then the question is: if I can store the data elsewhere, do I still need this API?

The host asked whether this would only be a matter within the geek community.

Peter said no. He had just attended an agent gathering in Vienna, where he met someone from a design company. This person had never written code before, but started using Clawdbot in December (before it became popular). Now, their company has 25 internal web services, all built by agents through Telegram conversations.

This is a paradigm shift. You no longer subscribe to random startups that only address 10% of your needs. You now have your own hyper-personalized software, precisely tailored to solve your problems—and it's free.

Also, don't forget that this is the worst possible time for models. They can only get better and faster from here.

The host asked him what he was going to do next.

Peter said he received a bunch of emails from security researchers.

The issue is that it was initially intended for personal use, with scenarios envisioned as one-on-one chats on WhatsApp or Telegram, with people you trust. Discord was added later, but the model's assumption is still that you trust the people in the group.

"Now people are using it in places I never imagined. That little web app which was originally for debugging purposes has been directly thrown onto the public internet. All the threat models I previously didn't care about are now popping up."

To be honest, this is all code I wrote based on vibes (vibe-coded). I just want to show a direction, not deliver an enterprise-level product. I'm not even sure any company would touch this, as there are still some unresolved issues. Prompt injection hasn't been resolved and definitely poses risks.

He said he included a warning on the website and in the onboarding process: "With great power comes great responsibility." Early users understood this, as many of them were AI researchers. But people who are joining now may not necessarily understand it.

"I think this will accelerate the research, because there is now a demand, and we must find ways to make it safe for everyone."

The host asked him if he would start a company.

Peter said he prefers foundations or nonprofit organizations.

"I haven't decided yet."

The host said, "Ten thousand VCs just made a hole in the wall."

Peter smiled.

The host asked him his opinion on open-source licenses, and whether someone might directly take the code and sell it.

Peter said it will definitely be the case.

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

He chose the MIT License.

Someone might take it and sell it, but it doesn't really matter. The code itself isn't valuable anymore. Even if you delete it, you can rebuild it in a few months. What's truly valuable is the idea, the attention, and the brand.

The host asked if there was anything else he wanted to say.

Peter said he needs help.

"If you love open source, have experience, enjoy handling security reports, or like reverse-engineering software but are also willing to help fix things—email me. I'm at my limit now."

"I hope this project will last longer than my life. I feel it's too cool to let it rot away."

The host asked him whether the previously mentioned unfinished project would still be released.

Peter said that it's more like a hobby. He has some ideas in his mind about "what this kind of thing could become," but he doesn't want to reveal too much.

Purely for the love of the game.

Peter Steinberger's story has a recurring theme: when big companies don't act, he takes action himself, and then it blows up.

This is how PSPDFKit is. This is how Clawdbot is as well.

If someone can create a product in 10 days by "coding by vibe" (vibe-coding) that causes a vertical rise in GitHub stars, then where exactly is the moat (competitive advantage)?

His answer: ideas, eyes, brand. And, do it well enough that others have no room to copy.

But a deeper question might be: when a personal assistant can actually help you order food, create fitness plans, SSH into your computer to wake you up, when it can check file headers, find API keys, and use curl to implement features you haven't even written yourself— are we really ready for that?

Peter himself said: the prompt injection issue hasn't been resolved, and the risk is real. This is code "thrown together by intuition," not an enterprise-level product.

But he also said: this is the worst time for the model. It will only get better.

For me, I'm still not comfortable entrusting these tasks to AI. 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?