One million published skills, nearly 280 million installs and more than 70 agent platforms loading them, from ChatGPT to Cursor, GitHub Copilot and VS Code. With those numbers behind it, Anthropic announced on October 22, 2026 that it is handing the Agent Skills standard to the Agentic AI Foundation (AAIF), the Linux Foundation project that already governs the Model Context Protocol. The SKILL.md format stops belonging to one vendor and starts belonging to everyone.
What exactly is being handed over
It is more than the spec. Anthropic is giving the AAIF the Agent Skills name and logo, the GitHub organization (github.com/agentskills/agentskills) and the agentskills.io domain where the specification lives. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. You do not need to join the foundation to contribute or to use it.
Anthropic co-founded the AAIF with Block and OpenAI and had already donated MCP to it, so this is the second move of the same strategy:
| MCP | Agent Skills | |
|---|---|---|
| What it solves | Connecting the agent to tools and data | Giving it packaged knowledge and workflows |
| Shape | Client-server protocol | A folder with a SKILL.md, scripts and references |
| Governance | Agentic AI Foundation | Agentic AI Foundation since October 22, 2026 |
| Stated adoption | — | 70+ platforms, 1 million skills, ~280 million installs |
Why it matters
Until now, writing a skill meant betting on a format controlled by a single lab, however open it was. Having the same SKILL.md work in Claude Code, Codex, Cursor or OpenCode depended on goodwill: everyone else wrote loaders for a standard they had no say in. With the donation, format changes go through a neutral process with rivals at the table, which is what a company needs before standardizing its own internal skills without fearing lock-in to one tool.
For anyone maintaining a skills repository, like the ones this blog reviews every day, the practical upside is that the format gets a stable owner and, if the roadmap holds, versions. As things stand, a skill written for one agent can load in another with subtle differences in how it is discovered or triggered; a versioned standard lets you declare which version you wrote against.
Winners and losers
Platform teams that want one internal skills catalog for all their agents come out ahead, and so do smaller tool vendors, who no longer compete with "Anthropic's format" but help shape it. Anthropic gives up control, but keeps something valuable: the ecosystem born in Claude Code becomes the one the whole industry adopts. The losers are the proprietary rules and instructions formats each editor has invented, which now have a harder time explaining why they are not just skills.
What we still don't know
The roadmap Anthropic mentions is a set of proposals, not commitments: numbered spec releases, a formal proposal process and security guidance for people who write and install skills. That last one is the most overdue. A skill can ship scripts the agent runs with the user's permissions, and there is still no standard for signing, declared permissions or provenance. With a million skills out there, that guidance is late; having a neutral foundation write it, rather than every directory on its own, is the best outcome it could get.
Sources
- Donating Agent Skills to the Agentic AI Foundation, Anthropic, October 22, 2026.
- Donating Agent Skills, Simon Willison, October 22, 2026.