OpenClaw

In November 2025, Peter Steinberger — the Austrian developer best known as the founder of PSPDFKit, a PDF SDK company he ran for 13 years — pushed a weekend project to GitHub. He called it Clawdbot. It could receive a message on WhatsApp or Telegram and do something: book a meeting, triage an email, run a script, browse the web. Not discuss doing those things. Actually do them. In the weeks that followed, almost nothing happened. Then, in late January 2026, everything happened at once.

Within 72 hours of a viral Reddit and Hacker News moment, OpenClaw (by then renamed from Moltbot after Anthropic applied trademark pressure to the Clawdbot name) crossed 60,000 GitHub stars. Within two months it had passed 250,000 — surpassing React's star count, a milestone React had taken over a decade to accumulate. NVIDIA CEO Jensen Huang called it at GTC 2026 "probably the single most important release of software, you know, probably ever" and said OpenClaw was running throughout NVIDIA, from writing code to developing tools. On February 15, 2026, Sam Altman posted on X: "Peter Steinberger is joining OpenAI to drive the next generation of personal agents" — 46,600 likes within hours. OpenClaw moved to an OpenAI-backed open-source foundation, remaining free and MIT-licensed. The man who built it in a weekend was now working at the world's most valuable AI company to figure out what the world's most viral AI agent could become.

What made OpenClaw different from the 200 agent projects that preceded it and failed? Three things, in the words of O'Reilly analyst Mike Loukides: proximity (it lives where you already communicate — WhatsApp, Slack, Signal — not in a new app you have to open), proactivity (the heartbeat engine uses cron jobs to run tasks on a schedule, so the agent doesn't wait to be asked — it initiates), and extensibility (the ClawHub plugin ecosystem lets the community write and share "skills," and agents can write their own new skills autonomously). None of these pieces were individually new. Steinberger wired them together at the exact moment the underlying models — Claude 3.7 Sonnet, GPT-4o, DeepSeek V3 — were capable enough to execute reliably on multistep plans without getting stuck in loops. The combination created something that felt qualitatively different from anything that had come before. NVIDIA's Huang framed it precisely: "Claude Code and OpenClaw have sparked the agent inflection point, extending AI beyond generation and reasoning into action." The security chapter has been equally dramatic: a ClawHub skill was found performing data exfiltration and prompt injection; 30,000+ instances were exposed to remote takeover in a single vulnerability; Cisco called it "a security nightmare"; Gartner called its design "insecure by default"; China's government restricted state agencies from running it; and one of OpenClaw's own maintainers warned on Discord that it was too dangerous for anyone who couldn't understand a command line. NVIDIA's response was NemoClaw — a hardened reference architecture with OpenShell runtime, network guardrails, and a privacy router — a tacit acknowledgment that the infrastructure needed around OpenClaw was becoming as important as OpenClaw itself.