Peter Steinberger's First Public Appearance Since Clawdbot's Viral Launch

A solo developer, back from three years of burnout-induced retirement, accidentally built what might be the fastest-growing open source project in GitHub history. Not by raising billions, not by assembling a team, but by scratching his own itch: he wanted to talk to his AI agent from WhatsApp while making coffee.

The Interview

Peter Steinberger joins TBPN hosts John Coogan and Jordi Hays from Vienna at 11 PM local time for his first public appearance since Clawdbot (now renamed Maltbot) went viral. The conversation covers his journey from iOS developer burnout to accidentally creating a global phenomenon, his philosophy on building for agents, and why he thinks 2026 is the year of the personal assistant.

The tone is remarkably casual for someone sitting on what multiple VCs are frantically trying to fund. Steinberger repeatedly emphasizes he built this for fun, doesn’t need money, and would prefer a foundation over a company.

From Burnout to “Claude Code Anonymous”

Steinberger ran his own software company (PSPDFKit) for 13 years in the Apple/iOS ecosystem, sold it about four years ago, then hit a wall. Three years of complete burnout followed, which he jokes about but clearly took seriously.

The math, as he puts it: for every four years of work, you need one year of recovery. Thirteen years nonstop meant three years off checked out.

In April 2025, his spark came back. Coming from iOS, he wanted to build web and cloud things but didn’t want to feel like an idiot. So he looked into AI-assisted coding. Having missed the years when it was “really bad,” he walked right into Claude Code’s beta release.

“I literally had trouble going to bed. We had addiction before and then we had addiction again.”

He started texting friends at 4 AM about Claude Code. They replied. He started a meetup called “Claude Code Anonymous” (now rebranded to “Agents Anonymous” to keep with the times). His Twitter bio: “Came back from retirement to mess with AI.”

The Moment It Clicked: Voice Messages in Marrakesh

The origin story of Clawdbot is pure accidental discovery. Steinberger hacked together a WhatsApp integration in about an hour: receive a message, call Claude Code, return the response. Simple one-shot relay.

Then he added image support, because screenshots give agents rich context without typing. On a birthday trip to Marrakesh, he found himself using this constantly, not for programming, but for finding restaurants, figuring things out on the go.

Then the pivotal moment: he instinctively sent a voice message, forgetting he hadn’t built voice support. The read indicator appeared. Ten seconds later, the agent replied as if nothing happened.

When he asked how, the agent explained its chain of reasoning: the message contained a link to a file with no extension. It inspected the file header, identified it as Opus audio. It used ffmpeg to convert it to WAV. It tried to use Whisper but found an installation error. It looked around, found an OpenAI key in the environment variables, sent the audio via curl to OpenAI’s API, got the transcription back, and responded.

“That was the moment where, wow. These things are damn smart, resourceful beasts if you actually give them the power.”

This is what convinced him there was something real here. Not the coding capabilities, but the agent’s ability to improvise with whatever tools were available on the system.

CLIs Beat MCPs: Building for How Models Think

Steinberger has a strong and specific technical opinion: MCPs (Model Context Protocol servers) don’t scale. CLIs do.

His argument: agents already know Unix. You can have a thousand small CLI programs on your computer. The agent just needs to know the name, call the help menu, understand the interface, and use it. If you build the CLI to output what models already expect (structured, predictable, machine-readable), everything works better.

“Don’t build it for humans, build it for models.”

Before building Clawdbot, he created dozens of CLIs: one for Google Places, one for his Sonos system, one for his security cameras, one for home automation. Each CLI gave his agent more power and made the whole system more useful. It’s a new kind of software, he argues: agentic-driven tools built for how models think rather than how humans navigate GUIs.

He built the “world’s most expensive alarm clock” by having his agent SSH into his London MacBook to turn up the volume. He gave the agent a heartbeat with the prompt “surprise me.” The whole approach is equal parts technology and art.

Why Apps Will Melt Away

Steinberger makes a specific prediction about the future of software: a whole layer of apps will disappear because people will interact differently with AI agents.

His example: why do you still need MyFitnessPal? Take a picture of your food, and the agent already knows you’re at McDonald’s making bad decisions. It has your fitness goals, your workout program, your dietary history. It adjusts everything automatically. No app needed.

“Most apps will be reduced to API, and then the question is, do you still need the API if I can just save it somewhere else?”

He met someone at a Vienna meetup, a design agency with zero coding experience, who discovered Clawdbot in December and now runs 25 internal web services built entirely through conversations with the agent. Non-technical people building custom software that solves exactly their problem, for free, because talking to an agent comes naturally.

The Rename Drama and Anthropic’s Email

One of the more colorful segments: Anthropic contacted Steinberger asking him to rename the project (originally called Claudebot, too close to their trademark). He gives them credit for being “really nice” about it, sending an internal person rather than lawyers.

The execution was chaotic. He had two browser windows open simultaneously. Pressed rename on one, and by the time he finished creating the new account on the other, crypto squatters had already snatched the handle with automated scripts. X’s team helped resolve it quickly, but for about 20 minutes things were messy.

The hosts note the new name (Maltbot) works well and having brand independence will be good long-term. Steinberger agrees but clearly found the timing stressful.

Security, Scale, and the Solo Developer Problem

With viral traction came problems Steinberger never designed for. The project was built for personal use: one person talking to their agent on WhatsApp or Telegram in a trusted environment. Now people are exposing the debugging web app to the open internet, creating threat models he never considered.

“What people don’t realize is this is not a company. This is one dude sitting at home having fun.”

He’s honest about being overwhelmed. At one point he was copy-pasting Discord questions into Codex for automated responses, then copying entire channels with “answer the 20 most asked questions.” Security researchers are bombarding him with reports, some valid for use cases he never intended to support.

His pragmatic take: prompt injection isn’t solved yet. No company would ship this as a product because fundamental problems remain unsolved. But the demand might be what accelerates the research to fix them.

Foundation Over Company, MIT Over Restrictions

When asked about forming a company (with VCs “frantically trying to give you money”), Steinberger says he’d much rather consider a foundation or nonprofit. The hosts joke that “10,000 VCs just punched a hole in the wall,” then note the irony that investing in nonprofits has actually worked out well recently.

On open source licensing, he’s equally philosophical. He chose MIT knowing people will fork it and sell it. His defense strategy: make the open source version so good there’s not much room for commercial forks to differentiate.

“Code is not worth that much anymore. You could delete it and build it again in months. It’s much more the idea and the eyeballs and maybe the brand that actually has value.”

On Models: Opus for Character, Codex for Reliability

A brief but revealing aside on model preferences. Steinberger rates Opus as the overall best, particularly for personality and Discord interactions. He programmed a “no reply” token so the agent doesn’t spam every message, just listens and occasionally drops a joke that actually lands.

For coding, though, he prefers Codex (OpenAI): “I literally prompt and push to main with 95% certainty it works.” Claude Code requires “more tricks, more charade” to achieve the same result, though both are good. He can parallelize faster with Codex because it requires less handholding.

He runs a maxed-out Mac Studio (not a Mac Mini; “my agent is a bit of a princess”) with 512GB to experiment with local models, naming MiniMax-21 as the best current open source model.

Afterthoughts

The interview captures a rare moment: a genuinely independent developer at the center of a cultural phenomenon, explicitly choosing not to capitalize on it in conventional ways. A few threads worth pulling on:

• The improvisation thesis is the real story. Not that AI can code, but that an agent with system access will figure out solutions the developer never anticipated. The Marrakesh voice message incident isn’t an anecdote; it’s a design philosophy.
• “Build for models, not humans” is a software architecture shift. If Steinberger is right that CLIs outperform MCPs for agent interaction, it inverts decades of UX thinking. The best interface for the next generation of software might be --help.
• The apps-melting-away prediction is already testable. A non-technical design agency running 25 custom web services through agent conversations is a data point, not speculation.
• The solo developer paradox. One person can now ship at the velocity of a small company, but the demands of virality (security, support, governance) still require organizational scale. Steinberger’s answer, a foundation, is an experiment in itself.