From Burnout to AI Agent Stardom: Peter Steinberger’s Moltbot Journey

1. A 13‑year startup, three years of burnout, then Claude Code appears

Peter Steinberger, founder of PSPDFKit, sold his company in 2021, suffered a severe burnout, and disappeared for three years. In November 2025 he spent ten days “vibe‑coding” a new AI‑driven tool that quickly became Clawdbot, later renamed Moltbot after a request from Anthropic.

“Last year was the year of programming agents; this year is the year of personal‑assistant agents. I think I’ve lit a fire.”

The interview (YouTube link) is his first public appearance after the project’s viral success.

2. 4 am messages – “We’re all addicted”

Peter admits he became addicted to AI programming, pulling friends into an “Agents Anonymous” group that meets at any hour.

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

His GitHub profile now reads: “Retired, back to playing with AI, having fun.”

3. Early ideas stalled by big companies

He had a personal‑assistant concept in May 2025, tried GPT‑4, found it unsatisfactory, and waited for large companies to release similar tools. Six months later, none had, so he built his own.

“Where is my damn agent?”

4. WhatsApp integration in one hour

Peter asked himself each morning what would be cool to build. The answer: chat with his computer via WhatsApp. In one hour he hacked together a basic version that receives WhatsApp messages, forwards them to Claude Code, and returns the result.

“If my agent is running and I’m in the kitchen, I want to check its status or send a quick command.”

He added image support because screenshots provide rich context for prompts.

5. Marrakech night – “If you give them real power…”

During a birthday weekend in Marrakech, Peter used the tool to look up restaurants and information, discovering its utility beyond coding.

“Because it integrates Google, it can fetch things on the fly, which is especially useful when you’re out.”

When he sent a voice message without any voice‑processing code, the agent detected the Opus file, converted it with ffmpeg, used Whisper via OpenAI’s API, and replied—all automatically.

6. The “most expensive alarm clock” and “surprise me”

Peter turned the agent into a personal alarm: the AI SSHes from a London server into his Vienna MacBook, raises the volume, and wakes him up.

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

He added a heartbeat that periodically sends a “surprise me” prompt, turning the system into a glue that stitches existing services together.

7. MCP is trash – CLI tools scale

Peter argues that Model Context Protocol (MCP) cannot scale, whereas command‑line tools can. By exposing a --help menu, agents can discover and invoke functionality without human‑centric design.

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

He built dozens of CLI wrappers for Google, Sonos, cameras, and home‑automation, giving the agent new capabilities.

8. 72‑hour explosion – Discord flood

When the project blew up, Discord grew at an unprecedented rate. Peter initially used Codex to answer individual questions, then batch‑processed the top 20 FAQs and posted the answers en masse.

“People don’t realize this isn’t a company; it’s a single person playing at home.”

9. Model evaluation – Opus vs. Codex

Peter’s playground supports any model. He finds Opus has a strong “personality” but Codex is more reliable for code generation.

“Codex handles large codebases better; 95 % of the time it can push directly to main and the code runs.”

10. Renaming drama – Anthropic’s request

Anthropic asked for a name change due to trademark issues. During the rename, a crypto scammer grabbed the old handles within ten seconds, launching a fake $CLAWD token that briefly hit a $16 M market cap before Peter publicly denied it.

11. Mac Studio vs. Mac Mini for local models

Peter runs his agents on a 512 GB‑RAM Mac Studio because current local models (e.g., Mini‑ax 21) still need multiple machines.

“Running locally bypasses the auth hurdles of cloud services.”

12. Many apps will disappear

He predicts most consumer apps will be reduced to simple APIs that an AI agent can call, eliminating the need for dedicated apps like MyFitnessPal.

“Most apps will become APIs; if I can store the data elsewhere, why keep the API?”

13. Security researchers flood in

After the launch, security researchers contacted him about prompt‑injection risks. Peter admits the code is “vibe‑coded” and not enterprise‑grade, and that prompt‑injection remains unsolved.

“The code is built on feeling; prompt‑injection is a real risk.”

14. Foundation, not a company

He leans toward forming a non‑profit foundation rather than a for‑profit company, using the MIT license to keep the project free and widely accessible.

“Code itself isn’t valuable; ideas, attention, and brand are.”

15. Recruiting maintainers

Peter openly asks experienced open‑source contributors, security‑report handlers, and software tinkerers to help maintain the project, hoping it outlives him.

“If you love open source and want to keep this alive, email me.”

The interview ends with a reflection: last year was the year of programming agents; this year may become the year of personal‑assistant agents, but the technology is still immature and carries real risks.