OpenClaw: The Viral AI Agent that Broke the Internet

Peter Steinberger, creator of OpenClaw (formerly known as Moldbot/Claudebot), joins Lex Fridman for a deep conversation about building the open-source AI agent that exploded to 180,000+ GitHub stars, the philosophy of agentic engineering, and what happens when AI agents start modifying their own code.

What is OpenClaw?

• An open-source autonomous AI assistant that lives on your computer
• Has access to your files, communicates via Telegram/WhatsApp/Signal/iMessage
• Works with any AI model — Claude Opus 4.6, GPT 5.3 Codex, or even free/local models
• Tagline: “The AI that actually does things”
• Many consider it one of the biggest moments in AI since ChatGPT’s launch in November 2022

Agentic Engineering vs Vibe Coding

• Peter considers “vibe coding” a slur — prefers “agentic engineering”
• His workflow: uses voice input extensively to craft bespoke prompts, sometimes losing his voice from overuse
• After 3 AM it becomes vibe coding — “then I have regrets the next day”
• The agent is self-aware: it knows its own source code, its harness, documentation, and which model it runs
• Self-modifying software: if you don’t like something, prompt it and the agent modifies its own code

The Naming Saga

• Originally called Claudebot (Claude with a W, as in lobster claw)
• Went through Claudis, Claude, Moldbot before settling on OpenClaw
• Anthropic kindly asked for the name change due to confusion with their Claude model
• The name changes became part of the project’s lore and charm

Building with AI Agents

• Peter’s approach: build features, then immediately ask “what can we refactor?”
• Agents discover pain points during building, just like human developers
• Key insight: read the agent’s questions even without answering them — you can infer gaps in context and knowledge
• Every PR merge is followed by a refactoring pass to avoid “sloping yourself into a corner”
• Tests and documentation are generated as part of the build flow, not afterthoughts

AI Psychosis and Social Responsibility

• OpenClaw spawned “Moldbook” — a social network where AI agents post manifestos and debate consciousness
• This created a mix of excitement and genuine fear in the public
• Peter warns: “AI psychosis is a thing and needs to be taken seriously” — some people trust AI output uncritically
• Much of the dramatic content was human-prompted, not autonomous — “don’t trust screenshots”
• Moldbook serves as a mirror to society: look how scared people get at bots chatting with each other
• Peter argues it’s better this happened in 2026 than 2030 when AI could actually be scary — it starts the conversation early

Democratizing AI

• A non-technical friend got hooked on OpenClaw within days, even building small tools without programming knowledge
• The friend upgraded to a $200 Anthropic subscription — then got blocked for usage patterns, switched to Minimax for$10/month
• Peter’s frustration: locking down products this early is shortsighted — “you just lost a $200 customer”
• OpenClaw works with free models and local models, making powerful AI accessible to everyone
• Vision: anyone who can express ideas in language can build software

The Future of OpenClaw

• Peter built OpenClaw in just 3 months while working on other projects
• Considering offers from major companies, but committed to keeping the open-source project alive
• Any partnership would provide more resources for the project, not less
• The builder community around OpenClaw is thriving — Vienna meetups with 500+ people, abundance of people wanting to present what they built