OpenClaw reached 200,000 GitHub stars in months by combining three architectural decisions: local-first data storage using plain Markdown/YAML, messaging apps (WhatsApp, Telegram, Slack) as the interface, and true model-agnosticism supporting Claude, GPT, DeepSeek, and Ollama. Its Skill System drove viral adoption through community-extensible workflows. However, Cisco research found critical vulnerabilities in third-party skills, with ~26% of 31,000 analyzed skills containing at least one flaw, exposing a fundamental tension: the same openness that built developer trust created serious security risks. Key lessons include prioritizing deployment topology and user sovereignty over model performance, focusing on the agent harness layer as the real long-term opportunity, and establishing governance frameworks before viral growth rather than after.
Table of contents
It's not about the model. It's about everything around it.It’s not the model – It’s the harnessThe architecture that workedExtensibility and the riskBuilding trust in the community and the security realityLessons learnedMore from We Love Open SourceAbout the AuthorSort: