Steinberger, Peter
▶ Watch on YouTubeCore ideas
Steinberger tells OpenClaw's origin as an accident of speed: a roughly one-hour prototype connecting WhatsApp to Claude Code, built mostly to scratch his own itch, that turned into a viral phenomenon he never designed for scale or fame. The "mind-blowing moment" that convinced him something real was happening was watching the agent handle an audio-transcription task he had never explicitly taught it — it figured out the right tool chain on its own, which crystallized for him that the interesting frontier isn't hand-built features but giving a capable model enough autonomy and the right primitives to solve novel problems itself. He argues the project's viral spread wasn't really about utility but about fun and weirdness — people were captivated by an agent that felt alive, self-modifying, and a little unpredictable — and he's candid that this same "aliveness" produced real chaos, most vividly the multi-day, multi-name saga (WA-Relay to Claude's to ClaudeBot to MoltBot to OpenClaw) in which a botched atomic rename briefly let crypto squatters seize accounts and domains, forcing him to improvise a rebrand under pressure in public view.
On engineering philosophy, Steinberger is skeptical of MCP as a long-term paradigm, arguing that a plain CLI is a more natural affordance for models than a bespoke structured protocol: an LLM already excels at calling Unix commands, and a CLI is composable (the model can pipe output through `jq` or chain commands to filter and shape context) in a way that a monolithic MCP response blob is not, since MCPs tend to dump everything into context whether the model needs it or not. He extends this to a broader worldview that any application "is just a slow API" if a browser-driving agent can operate it, predicting an accelerating shift where personal agents route around companies that don't build agent-friendly interfaces (his own reverse-engineered Twitter CLI, "Bird," and Google-focused CLI, "GAWK," are examples), while still voicing sympathy for the platforms' incentive to slow that access down. He is emphatic that only content or fully-human labor should be presented as human — he blocks any account that tweets via AI on sight — and describes developing a strong personal aversion ("AI slop... it's a smell") to AI-generated prose, images, and infographics, even material he made himself months earlier, arguing authenticity is becoming more, not less, valuable as machine-generated content becomes cheap and ubiquitous.
He describes his own coding workflow as voice-driven and deliberately un-precious: no reverting agent work, keeping main always shippable, and consciously empathizing with the agent's fresh, context-less perspective on each task rather than assuming it shares his accumulated intuition — a discipline he calls avoiding the "agentic trap." Drawing on his thirteen years building PSPDFKit (including a long stretch of burnout and a deliberate multi-year hiatus before returning to build OpenClaw), he frames money as secondary to the experience of building something meaningful, and turned down acquisition interest from major labs (OpenAI, Meta) on the condition that OpenClaw stay open source, drawing an explicit analogy to Chrome/Chromium's dual open/commercial model. On the bigger economic and social picture, he predicts a large fraction of consumer apps become obsolete as personal agents subsume their functions directly (fitness tracking, smart-home control, calendar apps), pushing companies to either become thin, agent-facing APIs or fade like Blockbuster; he sees this reshaping programming itself into something closer to orchestration and taste rather than hand-written code, and while he acknowledges real grief in that transition for people whose identity is bound up in the craft, he insists the deeper human skills — vision, architecture, empathy for the tool and the problem — remain scarce and valuable. He closes on hope drawn directly from user stories: small businesses automating tedious back-office work, a disabled person's daughter feeling newly empowered, and a groundswell of amateur builders (citing a Vienna meetup, "ClawCoin," packed with first-time presenters) that he reads as evidence the technology, net, gives ordinary people real agency rather than just displacing them.
Excerpts
On why he thinks CLIs beat MCP as an interface for models, and the composability problem with structured protocols:
“My approach is if you want to extend the model with more features, you just build a CLI and the model can call the CLI, probably gets it wrong, calls the help menu, and then on demand loads into the context what it needs to use the CLI... And MCP is... That has to be added in training. That's not a very natural thing for the model... And the biggest thing, it's not composable. So imagine if I have a service that gives me... a huge blob back. As a model, I always have to get the huge blob back... There's no way for the model to naturally filter unless I think about it proactively... But if I would build the same as a CLI... it could just add a JQ command and filter itself.”
On his zero-tolerance reaction to AI-generated content and why he thinks authenticity becomes more prized, not less:
“If you tweet at me with AI, I will block you. No first strike. As soon as it smells like AI, and AI still has a smell... I much rather read your broken English than your AI slop. You know, of course there's a human behind it, and yet they prompt it. I'd much rather read your prompt than what came out. I think we're reaching a point where I value typos again.”
On why he insisted any acquisition keep OpenClaw open source, and his broader philosophy of building for experience over money:
“Describing conversations with major labs interested in acquiring the project, he insisted it remain open source in a Chrome/Chromium-style model — commercial interests can build on top, but the core stays open — reflecting a stance he traces back through his PSPDFKit years and burnout: that the value of building something is in the experience and the people it reaches, not in maximizing the payout.”