Remote OpenClaw Blog
The Story Behind OpenClaw: From Side Project to Production AI Agent
8 min read ·
Remote OpenClaw Blog
8 min read ·
Every large project starts small. OpenClaw started as a few hundred lines of TypeScript written over a weekend in November 2025 by Peter Steinberger, an Austrian developer who had just stepped away from running PSPDFKit for 13 years.
The idea was simple: connect WhatsApp to Claude's API so he could text his AI assistant from his phone instead of opening a browser tab. No framework ambitions. No GitHub stargazing strategy. Just a developer scratching his own itch.
He called it ClawdBot — a play on "Claude" and "bot" — and pushed it to a public GitHub repository mostly as a backup. The README was two paragraphs. The architecture was a single file.
Within a week, other developers had found it, forked it, and started asking questions. That was the first sign that this wasn't going to stay small.
The choice to start with WhatsApp wasn't random. Steinberger lived in Austria, where WhatsApp is the default messaging app — used by over 90% of the population. His reasoning was pragmatic: he already had WhatsApp open all day, so making the AI accessible there meant he'd actually use it.
This turned out to be one of the most important early decisions, even though it seemed trivial at the time. By building around a messaging app rather than a web interface, OpenClaw was inherently designed for asynchronous, mobile-first interaction. You could text your agent while walking to the grocery store and get a response by the time you arrived.
The messaging-first approach also meant the agent had to be good at handling natural language instructions. There were no buttons, forms, or structured inputs — just text messages. This forced the underlying architecture to be flexible about how tasks were defined and executed.
Looking back, three architectural decisions made during the first month defined OpenClaw's trajectory. None of them were made with grand strategy in mind — they were pragmatic choices that happened to compound.
Steinberger ran ClawdBot on his own server from the start. He didn't build a hosted service or SaaS platform — he just wrote a Docker Compose file and published the setup instructions. This meant every early user was also a self-hoster, which created a community of people comfortable with servers and DevOps. These were exactly the kinds of contributors who could improve the codebase.
Self-hosting also meant data privacy was built into the DNA. Your conversations, your memories, your API keys — everything stayed on your hardware. In an era of increasing concern about AI companies training on user data, this was a powerful value proposition.
The original version only worked with Claude (Anthropic's model), but Steinberger designed the model interface as an abstraction layer rather than hardcoding Claude's API. When other developers asked for GPT support, adding it was a straightforward implementation of the existing interface.
By December 2025, ClawdBot worked with Claude, GPT-4, and local models via Ollama. This model-agnostic approach meant the project wasn't tied to any single AI provider's success or pricing changes. Operators could switch models without changing their workflows.
Rather than building every integration himself, Steinberger created a simple plugin system where anyone could add new capabilities by dropping a JavaScript file into a plugins directory. This was crude by modern standards — no type checking, no dependency management, no sandboxing — but it was enough to let the community build what they needed.
The plugin system turned OpenClaw from a single-developer project into a platform. By January 2026, there were plugins for Google Calendar, Notion, Trello, email, file management, and dozens of other tools. Each plugin made OpenClaw more useful, which attracted more users, who built more plugins.
By late January 2026, ClawdBot had outgrown its name. It wasn't a bot anymore — it was a framework for building autonomous agents. The name change to MoltBot reflected this evolution. "Molt" referenced the lobster mascot shedding its shell, a metaphor for the project's transformation from a simple relay into something much larger.
The MoltBot phase lasted only a few days, but it was significant because it coincided with two major additions: persistent memory (the agent could remember things across conversations) and scheduled tasks (the agent could do things on its own without being prompted). These features turned OpenClaw from a reactive chatbot into a proactive assistant.
Persistent memory, in particular, changed the user experience dramatically. An agent that remembers your preferences, your ongoing projects, and what you discussed last week is fundamentally different from one that starts fresh every conversation. For more on how memory works today, see who owns OpenClaw and the governance of data handling.
The growth curve was unlike anything typical of open-source projects. Most successful repos accumulate stars over years. OpenClaw went from a few hundred to 100,000 in roughly three months.
Several factors converged:
Marketplace
Free skills and AI personas for OpenClaw — browse the marketplace.
Browse the Marketplace →The viral growth also brought challenges. Issues piled up faster than Steinberger could triage them. The Discord server went from dozens to thousands of members in weeks. Feature requests came in faster than any single developer could evaluate them.
On January 30, 2026, the project was renamed to OpenClaw. This was the final name, and it stuck for good reasons: "Open" emphasized the open-source commitment, and "Claw" tied back to the lobster mascot that had become the project's visual identity.
The rename coincided with a major restructuring of the codebase. The single-file architecture was replaced with a modular system. Plugins were formalized. Configuration moved from hardcoded values to YAML files. Documentation went from a sparse README to a full docs site.
This was also when ClawHub launched — a public registry for community-built skills and integrations. Think of it like npm for OpenClaw plugins. Developers could publish their work, and operators could install it with a single command. Within weeks, ClawHub had over a thousand entries. To see how OpenClaw's open-source model works, check our dedicated guide.
In February 2026, Steinberger announced two things simultaneously: he was joining OpenAI, and OpenClaw would transition to a foundation.
The foundation model was designed to prevent a common failure mode in open source: a project dying because its sole maintainer moves on. By establishing a governance structure with multiple maintainers, a public roadmap process, and financial sponsorship (including from OpenAI), Steinberger ensured OpenClaw would survive and thrive without him.
The full story of who made OpenClaw covers the foundation transition in detail. The key point for this article: the foundation worked. OpenClaw continued shipping regular releases, the community kept growing, and the project's direction shifted from one person's vision to a collectively governed roadmap.
The OpenClaw story offers several lessons that apply beyond AI agent development:
For operators exploring OpenClaw today, browse the marketplace for production-ready personas and skills that build on the foundations described in this article.
OpenClaw started in November 2025 as a personal project by Peter Steinberger called ClawdBot. It was a simple WhatsApp relay that connected a phone to Claude's API. Steinberger kept adding features — persistent memory, tool calling, scheduling — and the project snowballed into a full AI agent framework that hit 100K GitHub stars in three months.
The most impactful decision was making OpenClaw self-hosted and model-agnostic from the start. By letting operators run their own instance with any AI model provider, OpenClaw avoided vendor lock-in and attracted developers who were privacy-conscious or wanted control over their AI stack. The open-source approach also enabled a community that built thousands of plugins and skills.
The project started as ClawdBot (November 2025), was renamed MoltBot (late January 2026), and became OpenClaw (January 30, 2026). Each name change reflected the project's rapid evolution — from a simple chatbot relay to a full agent framework to an open-source platform. The lobster mascot remained consistent throughout all three names. See who owns OpenClaw for more on the governance evolution.
Yes. When Peter Steinberger joined OpenAI in February 2026, he transitioned OpenClaw to a foundation model with community governance. The project is actively maintained by a core team of contributors, releases new versions every 2-3 weeks, and has a growing ecosystem of over 15,000 community-built skills on ClawHub.