There is probably no one in the AI community who hasn’t heard of OpenClaw – a project that quickly became one of the fastest-growing repositories on GitHub, surpassed 200,000 stars, and sparked a wave of new experiments and similar local agent projects.

In a matter of days, OpenClaw’s founder, Peter Steinberger, was interviewed by Lex Fridman (over 600k views) and hired by OpenAI (who even knows for how many millions) – a sign that this was more than just another incremental agent project.

What distinguished OpenClaw was its relatively light safety scaffolding, which made the underlying capabilities more visible and easier to experiment with. In that sense, Steinberger’s trajectory loosely echoes OpenAI’s early ChatGPT moment: a decision to release something imperfect but powerful, and let the consequences reshape the conversation. And the whole world of personal assistants.

And what a blunder by Anthropic. Initially, the project was called Clawdbot – a literal homophone of Claude. Instead of embracing the momentum around it – by backing the project, hiring Peter, or integrating it into their coding-agent ecosystem – they sent lawyers and forced a rename. Facepalm.

In any case, today, we want to dedicate some time to unpack why OpenClaw matters and how it works under the hood, because the architecture shapes what this system can and cannot do. There are a few very clever ideas. We’ll also suggest a few lightweight alternatives and cut through the noise to understand what the OpenClaw debate is really about.

Whether you’ve played with it or not, there’s plenty to discover. Let’s go!

In today’s episode, we will cover:

The Boom

OpenClaw’s rise did not come from a massive corporate launch or a carefully staged rollout. As Peter Steinberger describes it, he built the initial prototype – a WhatsApp relay connected to Anthropic’s Claude – in about an hour. And that revealed that instead of switching between apps, you could simply message an agent and have it act on your behalf.

Peter could never have imagined that this small experiment would soon disrupt the entire AI industry.

This marks a broader shift in how we interact with software. Peter Steinberger points out that many apps exist mainly to gatekeep interfaces. Once an agent has unified context – your location, sleep, calendar, preferences – it can reason and act across domains in ways no siloed app ever could.

Some apps will shrink into backend APIs; others will simply disappear. OpenClaw is the clearest embodiment of the practical, context-aware automation people have wanted for years.

The project has unleashed a surge of builder energy. California’s scene went through the roof: San Francisco’s Claw’s event drew 700+ developers, with follow-up meetups in the Bay Area and LA still selling out overnight and turning away hundreds. Vienna’s ClawCon packed 500 people who came to demo what they’d built. As Peter said: “The lobster is taking over the world.”

Beyond the tech itself, Peter’s personal story has become a major source of inspiration. After more than a decade scaling PSPDFKit, a commercial toolkit for working with PDF documents inside apps, he hit burnout and lost the joy of coding. He stepped back, traveled, and only rediscovered the fun through hands-on experiments with AI agents. Because he was financially secure, he could afford to run OpenClaw at a personal infrastructure cost of $10–20K per month — a tradeoff he was okay with because he believed deeply in the direction of personal agents. That mix of playful experimentation and genuine generosity, without the usual money-first pressure, is exactly what many builders say has been missing.

It all lands perfectly in this new era of open exploration and the renewed power of open source.

And then another big thing happened: Peter was hired by OpenAI while moving OpenClaw to an independent open-source foundation that OpenAI will continue to support. This is far more than a big company hiring talent. (Stay tuned to the end to see exactly why it matters so much.)

How is OpenClaw built? Architectural breakdown

The single biggest design decision in OpenClaw is this: it treats the agent itself as a collection of files on disk rather than as code you have to write or prompts you have to keep re-injecting.
Identity, memory, skills, heartbeat rules, tool policies – everything lives in a workspace directory as ordinary Markdown files and folders. That one shift turns the agent from a disposable script into durable, versionable, inspectable infrastructure.

Since in every AI101 episode we break down the behind-the-scenes stuff and how AI actually works, that’s where we’ll kick things off with OpenClaw explanation →