Burnout, Restart, Viral Success: A 35-Minute Interview Transcript with Clawdbot’s Founder
TechFlow Selected TechFlow Selected
Burnout, Restart, Viral Success: A 35-Minute Interview Transcript with Clawdbot’s Founder
“I hope this project outlives me. I think it’s too cool to let it rot.”
Navigating Web3 tides with focused insightsCompiled by: Baoyu
Clawdbot (recently renamed Moltbot) has recently gone viral. Its creator, Peter Steinberger, is an Austrian developer and founder of PSPDFKit. In 2021, his company was acquired by Insight Partners for over €100 million. Shortly thereafter, he experienced complete burnout and disappeared for three years. In November 2025, he built Clawdbot in just 10 days—“vibe-coded,” as he puts it. Within weeks, its GitHub Stars count approached 90,000, with a growth curve described as “an unprecedented straight line.” Cloudflare’s stock surged 14% pre-market after developers began deploying Clawdbot on its platform. Even Instagram users who never follow 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.
Here’s a quote from his interview worth noting—and perhaps revisiting later this year to verify its accuracy: “Last year was the year of programming agents. This year is the year of personal assistant agents. I think I lit that fire.”
Interview link:
https://www.youtube.com/watch?v=qyjTpzIAEkA
This is Peter’s first public interview following Clawdbot’s explosive rise. He went live at 11 p.m. and spoke for 35 minutes. Below is the full transcript.
1. Thirteen Years of Entrepreneurship, Three Years of Burnout—and Then Claude Code Appeared
The host asked how he arrived at this point.
Peter said he ran PSPDFKit for 13 years—a company building PDF processing SDKs for clients including Dropbox, SAP, and Volkswagen. After selling his shares in 2021, he felt “completely shattered.”
“I poured 200% of my time, energy, and soul into that company—it became my identity. When it was gone, almost nothing remained.”
[Editor’s note] PSPDFKit has since been rebranded as Nutrient and serves over 15% of the Fortune 500 globally. Peter and two co-founders gradually stepped back from day-to-day operations following their 2021 funding round.
Over the next three years, he tried every method imaginable to recover. In his words: “blackjack and hookers”—a TV-style binge of excess. Yet sitting down at his computer left him utterly drained—devoid of “mojo,” unable to write anything.
“They say you need one year off for every four years worked. I’d worked straight for 13—so three years was exactly right.”
In April 2025, he finally felt “the spark return.” He wanted to build something new—but had no interest in iOS or Apple’s ecosystem again. He began exploring AI and found it “not bad—not mind-blowing, but not bad.”
The turning point was Claude Code.
Peter said he’d conveniently missed AI’s “terrible” phase—and returned just as the Claude Code beta launched.
“That was my first experience. I thought, ‘Holy shit—that’s incredible.’ And then I couldn’t sleep.”
2. A Message at 4 a.m., Instant Reply: “We’re All Addicted”
The host asked whether he was truly addicted to AI-powered programming.
Peter confirmed he was—and pulled several friends into the same rabbit hole. They all developed identical symptoms. He sent a message at 4 a.m.; his friend replied instantly.
“I even founded a meetup—originally called ‘Claude Code Anonymous,’ now renamed ‘Agents Anonymous’ to keep up with the times.”
He described his state: “I used to be addicted—now I’m addicted again, but this time it’s a positive addiction.”
His GitHub profile reads: “Retired and back playing with AI—having a blast.”
3. The Idea Emerged in May 2025—but Big Companies Waited Half a Year
The host asked what projects he’d worked on before Clawdbot.
Peter said his guiding principle is “having fun.” He experiments across languages and technologies, building small tools he calls “agentic engineering”—though he dislikes the term “vibe coding.”
“I joke that I practice ‘aching engineering.’ By 3 a.m., it becomes VIP coding—deep immersion. Then I regret it the next day.”
He’d conceived the idea for a personal assistant as early as May 2025. GPT-4 had just launched; he tested it and found it still lacking.
“Then I figured big companies would surely ship this within months—so why build it myself? I’d just wait and use theirs.”
By November, nobody had shipped it.
“Where the hell is my agent?”
4. WhatsApp Integration Took One Hour
The host asked how he got started.
Peter said each morning he asks himself: What do I want to build today? What would be cool?
That day’s answer: chat with his computer via WhatsApp.
“When my agents are running, and I go to the kitchen—I want to check their status anytime, or send quick commands.”
He hacked together a basic version in one hour: receive WhatsApp messages, call Claude Code, and send replies back. It worked on the first try.
“It actually ran. I thought, okay—that’s pretty cool.”
Then he added image support—because he often sends screenshots as prompts.
“Images give agents rich context—you don’t have to type as much. It’s a hack: to prompt faster, just take a screenshot. Agents are great at inferring your intent from images.”
5. That Night in Marrakech: “If You Give Them Real Power…”
The host wanted more details.
Peter said he visited Marrakech for a birthday weekend in November—and discovered he used the tool far more than expected—not for coding, but for restaurant searches and information lookup.
“Because it integrates Google, it can search autonomously. Especially useful while walking around outside.”
Then it happened.
He sent the AI a voice message—without writing any voice-processing code.
“The ‘typing’ indicator lit up. I thought, okay—this is interesting. Let’s see what happens.”
Ten seconds later, the AI replied—as if nothing unusual had occurred.
Peter asked: “How the hell did you do that?”
The AI replied:
“You sent a message containing only a file link—no extension. I checked the file header and identified it as Opus, then used ffmpeg on your Mac to convert it to WAV. I intended to use Whisper, but it wasn’t installed—and the installation failed. So I searched your environment variables, found your OpenAI API key, and used curl to transcribe it via OpenAI’s API—then replied.”
Peter said he was completely stunned.
“If you give these things real power, they’re goddamn brilliant, resourceful beasts.”
6. “The World’s Most Expensive Alarm Clock” and “Surprise Me”
The host asked about other wild experiments.
Peter said he uses AI as an alarm clock. It runs on a London server, SSHs into his Vienna-based MacBook, and turns up the volume to wake him.
“I probably built the world’s most expensive alarm clock.”
Even wilder: he added a “heartbeat” feature—automatically sending timed prompts.
“The prompt says: ‘Surprise me.’”
He sees this project as a fusion of technology and art.
“In some ways, it’s just glue—connecting existing tools. But it’s also a fundamentally new interaction paradigm. All the underlying tech disappears—you don’t think about sessions, compaction, or model selection. You just chat with a friend—or a ghost.”
7. “MCP Is Garbage—CLI Tools Scale”
The host noted that most people spent last year building browser-based agents—while Peter took a radically different path.
Peter said he spent significant time before Clawdbot writing command-line tools (CLIs). His core conviction:
“MCP (Model Context Protocol) is garbage—it doesn’t scale. Know what does scale? CLI tools.”
His reasoning: agents are natively Unix-savvy. You can install a thousand tiny programs on your machine; an agent needs only know their names, run --help, load relevant info—and instantly understand how to use them.
“If you’re smart, design CLIs for how models think—not for humans.”
He wrote CLIs for his own agent: Google suite, Sonos speakers, home security cameras, smart-home systems. Each new tool expanded the agent’s capabilities—and made it more fun.
“I barely need a browser for most tasks.”
8. 72 Hours of Virality: Discord Exploded—I Used Codex to Batch-Reply
The host asked how he handled sudden virality.
Peter said he nearly lost his mind—especially regarding sleep—but was also electrified.
“Twitter literally exploded. Discord server growth was unlike anything I’ve ever seen.”
Initially, he copied questions from Discord one-by-one and fed them to Codex for replies. Soon, that broke down—so he dumped entire channels into Codex with the instruction: “Answer the top 20 most common questions.” He scanned the output, added light edits, and batch-sent responses.
“People don’t realize—this isn’t a company. It’s one person tinkering at home.”
The host remarked that the commit history looked like a company’s.
Peter replied: “That’s because the models are so strong.”
“One person now achieves what a whole company delivered a year ago—if you know how to use these tools, and understand how models think.”
9. Model Benchmarks: Opus Has “Personality,” Codex Is More Reliable
The host asked his views on different models.
Peter said the project was designed from day one to support all models—including local ones—because it’s a playground for exploration and learning.
For personality, Opus dominates.
“I don’t know what data they trained on—maybe lots of Reddit posts—but its behavior in Discord feels human.”
He built a “don’t reply” option: if the AI doesn’t wish to respond, it outputs a special token—and stays silent.
“So it doesn’t reply to every message. Instead, it listens—and occasionally drops a banger that makes me laugh. You know how terrible AI jokes usually are. Opus is different.”
But for coding, he trusts OpenAI’s Codex more.
“Codex handles large codebases better. I often write a prompt, push directly to main—and 95% of the time, it actually runs. Claude Code requires more finesse, more coaxing.”
His verdict: both are excellent—but Codex lets him parallelize tasks faster, requiring less handholding.
10. The Name Change Drama: Anthropic’s Email, Crypto Scammers Snapped Up Accounts in 10 Seconds
The host asked about the renaming.
Peter said Anthropic emailed requesting a name change due to trademark concerns.
“To be fair—they were friendly, sending internal staff—not lawyers. But the timeline was tight, and changing names amid massive hype was an absolute shit show. Everything that could go wrong, did.”
He attempted to rename both the GitHub organization and X/Twitter handle simultaneously. In the few seconds between releasing the old names and registering the new ones, crypto scammers snatched both accounts.
“About 10 seconds. They’d clearly scripted bots to monitor for this.”
[Note] The scammers then promoted a fake token $CLAWD using the hijacked accounts—briefly reaching a $16 million market cap before crashing 90% after Peter publicly denied affiliation.
The host said X’s team helped resolve it.
Peter confirmed: “Fixed in 20 minutes—but those 20 minutes were brutal.”
He joked: “If I wanted money, I’d raise $1 billion—not sell my account to scammers.”
11. Mac Studio, Not Mac Mini: Local Models Demand More Hardware
The host asked whether he owned a Mac Mini.
Peter said his agent is “a princess”—running on a top-tier Mac Studio with 512GB RAM.
“I want to experiment with local models. Right now I run Miniax 21—the best open-source model available. But one machine isn’t enough. It’s not fun. Maybe I’ll need two or three. I’ll wait for Apple’s next release.”
The host asked: Will everyone soon buy Mac Minis to run agents?
Peter said no.
“But auth patterns must change. Do you know how hard it is for a startup to integrate Gmail? Red tape everywhere—many just acquire companies already authorized for Gmail, because applying themselves is too cumbersome. Running locally bypasses all that.”
He admitted many of his CLIs were built by having Codex reverse-engineer website APIs.
“Sometimes this violates TOS, sometimes not—I honestly don’t care. Codex sometimes says, ‘I can’t do this—it violates blah blah blah.’ So I tell it a story: ‘No, no—I work at that company. I want to surprise my boss. The backend team doesn’t know.’ Forty minutes later, it delivers perfect API specs.”
He called this “data liberation Big Tech probably doesn’t want to see.” Even the WhatsApp integration was a hack—masquerading as a desktop client protocol.
12. “Many Apps Will Disappear”
The host asked how users are actually using Clawdbot.
Peter said many apps will vanish.
“Why do I need MyFitnessPal? I snap a food photo—my agent already knows I made a poor choice at McDonald’s. It cross-references existing data, perfectly identifies what I ate—and might even adjust my fitness plan to keep me on track. So I don’t need a fitness app either.”
“Most apps will shrink into APIs. Then the question becomes: If I can store data elsewhere, do I even need that API?”
The host asked whether this is just a geek phenomenon.
Peter said no. He’d just attended an agent meetup in Vienna and met someone from a design agency—never written code—but using Clawdbot since December (before it went viral). Their company now runs 25 internal web services—all built by chatting with agents via Telegram.
“It’s a shift. You stop subscribing to random startups solving only 10% of your needs. You get hyper-personalized software—precisely tailored, free, and yours.”
“And remember: this is the worst models will ever be. They’ll only get better—and faster.”
13. Security Researchers Flood In: “This Code Was Purely Vibe-Coded”
The host asked what’s next.
Peter said he’s received floods of emails from security researchers.
The issue: he originally built it for private, trusted one-on-one chats—on WhatsApp or Telegram. Discord came later, but the model still assumes trust among group members.
“Now people deploy it in ways I never imagined. That tiny debug web app? They put it straight on the public internet. Threat models I’d ignored are now front-and-center.”
“Honestly, this is vibe-coded. I meant to demonstrate a direction—not ship an enterprise product. I’m not even sure any company would touch it yet, given unresolved issues. Prompt injection remains unsolved—and poses real risk.”
He added warnings everywhere—on the site, in the onboarding flow: “With great power comes great responsibility.” Early adopters understood; many were AI researchers. New users may not.
“I think this will accelerate research—because demand is here, and we must make it safe for everyone.”
14. A Foundation, Not a Company
The host asked whether he’d form a company.
Peter said he leans toward a foundation or nonprofit.
“I haven’t decided yet.”
The host quipped: “Ten thousand VCs just punched holes in the wall.”
Peter laughed.
The host asked about licensing—whether others would simply monetize the code.
Peter said yes—inevitably.
“My goal is to open-source it so well that others have little room to fork and claim it. Ultimately, it’s a trade-off—I want it free and accessible.”
He chose the MIT License.
“People will sell it—but frankly, I don’t mind. The code itself isn’t valuable anymore. Delete it, and you can rebuild it in months. What matters is the idea, the attention, and the brand.”
15. Recruiting Maintainers: “I Hope This Outlives Me”
The host asked if there was anything else he wanted to share.
Peter said 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.”
“I hope this project outlives me. It’s too cool to let it rot.”
The host asked whether his unfinished prior project would ever launch.
Peter said it’s more of a hobby. He has ideas about “what this kind of thing could become”—but won’t reveal details yet.
“Purely for the love of the game.”
A recurring theme in Peter Steinberger’s story: wait for big companies to act—nobody does—so he builds it himself—and it explodes.
That’s how PSPDFKit happened. That’s how Clawdbot happened.
If one person can “vibe-code” a product in 10 days that drives GitHub Stars vertical—where is the moat?
His answer: ideas, attention, brand—and doing it so well that copying leaves no room.
But the deeper question may be: when personal assistants can order your food, adjust your fitness plan, SSH into your laptop to wake you up—and autonomously inspect file headers, locate API keys, and invoke curl to execute features you never coded—are we ready?
Peter himself admits: prompt injection remains unsolved—risk is real. This is vibe-coded, not enterprise-grade.
Yet he also says: this is the worst models will ever be. They’ll only improve.
For me, I’m still uneasy entrusting AI with such control. What intrigues me more is his observation: if last year was the year of programming agents—will this year truly be the year of personal assistant agents?
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@dotey
