Agent Skills work great in Claude Code. Too bad they're stuck there.
Skrun fixes that. It's an open-source CLI tool that takes the SKILL.md format Anthropic developed for giving AI agents specific capabilities and turns them into REST APIs. You write a skill once, run `skrun deploy`, and get a POST /run endpoint that anything can call. The developer built it because they wanted to use Agent Skills outside Claude Code, where the format currently lives. Skills locked inside one environment aren't much use if you're building on OpenAI or Mistral.
The tool supports five model backends (Anthropic, OpenAI, Google, Mistral, Groq) with automatic fallback. If one provider chokes, it shifts to another. Agents are stateful, remembering information across runs through a key-value store. That matters for something like an SEO audit agent that needs to compare results over time. Tool calling works two ways: CLI scripts you bundle with the agent, or MCP servers from the existing ecosystem. Demo agents cover code review and web scraping with Playwright, plus data analysis with typed input/output.
Skrun is early. About 10 days old on npm, pulling roughly 16 weekly downloads.
The current version only runs locally, though the codebase has a RuntimeAdapter interface ready for cloud deployment. There's also a registry for uploading and downloading agents, which suggests the founder is thinking about a marketplace angle. Commenters on Hacker News expressed interest in using it to test skills they've accumulated without being locked into Claude Code specifically.
Whether Skrun catches on depends on whether Agent Skills become a widely adopted standard or just another format that fragments across vendors. But the core idea is sound: agent capabilities should be portable, callable, and not trapped in one company's app.