OpenClaw: From Weekend Project to Most-Starred Repo on GitHub in 100 Days
AIopenclawai-agentopen-sourcelocal-firstself-hostedpeter-steinberger
Loading…
Back openDesk Edu for a sovereign, open-source education — every vote counts.
Vote nowOn 3 March 2026, OpenClaw surpassed React to become the most-starred software project on GitHub. React took 13 years to reach 243,000 stars. OpenClaw did it in about 100 days. The repository now sits at roughly 358,000 stars, 72,000 forks, 360 contributors, and 94 releases, with the latest being v2026.4.15. Nothing in the history of open-source software has grown this fast, not even close.
The story starts in November 2025 with Peter Steinberger, an Austrian developer best known for founding PSPDFKit, a document rendering SDK that powered over a billion devices and was acquired by Insight Partners. After the sale, Steinberger stepped away from coding for three years. When he returned in late 2025, it was to experiment with AI agents, and OpenClaw was his 44th attempt at an AI-related side project.
On a Friday evening, he sat down and built what he called "Clawdbot" in about an hour. It was simple glue code connecting WhatsApp with Claude Code, Anthropic's agentic coding tool. The idea was modest: instead of opening a separate browser tab to talk to an AI, you could just message it through the apps you already use. He published it on GitHub with an MIT licence and went to bed.
He woke up to roughly 800 Discord messages. The project had attracted 9,000 stars in its first 24 hours.
The name "Clawdbot" was a pun on "Claude" and the lobster mascot from Claude Code. It did not last. In late January 2026, Anthropic's legal team asked Steinberger to drop the Claude-adjacent branding. He renamed the project to Moltbot, a name chosen during what he described as a "chaotic 5am Discord brainstorm." The lobster theme stayed because, as Steinberger put it, "molting is what lobsters do to grow."
Three days later, on 30 January, it became OpenClaw. This time the naming was deliberate: proper trademark searches, domain purchases, and a name that reflected the project's open-source identity. The mascot, Molty the space lobster, remained throughout.
OpenClaw is written in TypeScript (89.9% of the codebase), with smaller contributions in Swift, JavaScript, Kotlin, Shell, CSS, and Go. At its core is a local-first Gateway that acts as a control plane. One long-running process on your machine, one WebSocket on localhost port 18789. The Gateway owns all messaging surfaces and brokers every message between channels, the AI agent, and local tools.
That gateway-first design is the key architectural decision. The AI model is a plugin to the gateway, not the other way around. Swap out Anthropic for OpenAI for a local Ollama model, and the gateway routing stays identical. Channels are isolated adapters, each normalising platform-specific messages into a standard internal format before passing them to the agent runner.
OpenClaw connects to the messaging platforms you already use: WhatsApp, Telegram, Slack, Discord, Google Chat, Signal, iMessage, Microsoft Teams, Matrix, IRC, WeChat, LINE, Mattermost, and over a dozen more. It supports voice on macOS, iOS, and Android through wake words and continuous talk modes, and offers a Live Canvas feature for agent-driven visual workspaces. Model support is provider-agnostic: bring your own API key or OAuth subscription for OpenAI, Anthropic, Google, or dozens of other providers, or run local models through Ollama, NVIDIA NIM, or any OpenAI-compatible endpoint.
The growth curve defies every pattern in open-source history. OpenClaw crossed 100,000 stars within days of going viral in late January 2026. It hit 145,000 by 11 February, 200,000 by mid-February, and 250,000 by early March when it overtook React. As of mid-April 2026, the repository has roughly 358,000 stars, 72,000 forks, and 360 contributors. The peak single day was 26 January, with over 25,000 new stars in 24 hours, a record for GitHub.
For context, Hugging Face took roughly four years to reach 100,000 stars. LangChain took 18 months. OpenClaw did it in under two weeks.
On 14 February 2026, Steinberger announced he was joining OpenAI. Sam Altman said Steinberger would "drive the next generation of personal agents." The project itself was moved to an independent open-source foundation, with OpenAI committing ongoing support. Reports indicated that both Meta and Anthropic had also pursued Steinberger, and that he had been spending roughly $10,000 per month of his own money on server costs for the project's infrastructure.
"What I want is to change the world, not build a large company," Steinberger wrote at the time, "and teaming up with OpenAI is the fastest way to bring this to everyone."
The explosive growth brought equally explosive security scrutiny. In February 2026, Microsoft Defender Research published a detailed analysis recommending that OpenClaw be treated as "untrusted code execution with persistent credentials." The core concern is straightforward: a self-hosted agent that processes untrusted text while holding durable credentials creates a dual supply-chain risk. Malicious skills from public registries can execute arbitrary code, and indirect prompt injection through external web content can steer agent behaviour without the user's knowledge.
Multiple CVEs were disclosed in early 2026, including a command injection flaw in the Docker sandbox (CVE-2026-24763) and a WebSocket authentication bypass (CVE-2026-28472). Bitdefender reported discovering over 800 malicious skills in the ClawHub registry. Security researchers scanning the public internet found thousands of exposed OpenClaw instances with minimal authentication.
The self-hosted model cuts both ways. You control the API keys and the data never leaves your device by default, which is a genuine privacy advantage over cloud-hosted alternatives. But you are also responsible for securing the gateway, hardening the runtime environment, and vetting every skill you install. Microsoft's recommendation was blunt: do not run OpenClaw on a standard workstation. Use a dedicated, isolated environment with tightly scoped credentials.
OpenClaw is not a coding agent in the mould of Claude Code or GitHub Copilot. It is a personal AI assistant that lives across your messaging channels, runs 24/7 on hardware you own, and performs real tasks: managing calendars, clearing inboxes, running terminal commands, booking flights, and automating workflows. The community calls it the "lobster way": always-on, local, fast.
That distinction matters. Coding agents sit inside an editor and help you write software. OpenClaw sits inside your messaging apps and helps you run your life. The difference is not incremental. It reflects a broader shift from AI as a tool you open to AI as a presence that is always there.
The project has spawned an entire ecosystem: language rewrites in Rust, Go, Python, and Shell; security tools like SecureClaw; deployment platforms; and Moltbook, an AI social network where agents interact autonomously. Nvidia CEO Jensen Huang called OpenClaw "definitely the next ChatGPT" at GTC 2026. Whether that prediction holds, the project has already changed the conversation around what personal AI agents can look like when they are open, local, and built by a community rather than a corporation.
Whether OpenClaw becomes a permanent fixture or a viral moment depends on whether the architecture can sustain the momentum. A hobby project that became infrastructure overnight now faces the challenge that all overnight infrastructure faces: security hardening, governance, and the messy work of turning enthusiasm into durability. The lobster has already exceeded every reasonable expectation. What happens next is less certain.