What is changing hands
Anthropic said on Thursday that it is donating Agent Skills, the open standard it uses to give AI agents procedural knowledge, to the Agentic AI Foundation, a project hosted by the Linux Foundation. The donation covers more than the specification. Anthropic is also handing over the Agent Skills name and logo, the GitHub organisation behind it and the agentskills.io domain.
It is the second standard Anthropic has given to the same foundation. It donated the Model Context Protocol there earlier, and it co-founded the foundation last year with Block and OpenAI.
A skill, in this standard, is a folder. It holds a file called SKILL.md that teaches an agent how to do one task or work in one domain, and it can carry scripts and reference material alongside. Anthropic said it built the format for Claude Code after finding that a model which handled general knowledge work well still lacked the procedural knowledge a real business task needs: the order of the steps, the house conventions, the things everyone on a team already knows.
The adoption numbers are Anthropic’s own
Anthropic said more than 70 agent platforms have integrated Agent Skills since the standard was published in December 2025, and it named ChatGPT, Cursor, GitHub Copilot and VS Code among them. It put the number of published skills at about a million and installations at nearly 280 million. Those are the company’s own figures, and the post does not say how either was counted.
The platform list is the part worth noticing. A standard that OpenAI’s own product reads is not something Anthropic can change on its own any more, and the donation formalises what the maintainer list already reflected: engineers from Anthropic, Google, OpenAI and other companies maintain the specification as individuals rather than as company representatives. Anthropic named David Soria Parra and Mahesh Murag as Lead Maintainers.
Security guidance is on the roadmap, not in the standard
The proposed roadmap under the foundation has three items: versioned releases, a formal proposal process for changes, and security guidance for people who write and install skills.
The third matters most and exists least. A skill is executable context, a folder of instructions and optionally scripts that an agent reads and acts on. Installing one from a registry is closer to installing a package than to pasting a prompt, and the standard as published does not tell either side what to check. With roughly 280 million installations already counted, the guidance is arriving after the practice rather than before it.
What this does not change
Anthropic said the standard stays open and vendor-neutral, and that nobody has to join the foundation to use Agent Skills or to contribute to it. Feedback on specification changes runs through the project’s Discord, and discussion happens on GitHub.
What to watch is the first versioned release. Governance moves like this are easy to announce and slow to show. The test is whether a change proposed by a maintainer outside Anthropic reaches a release, and how long the new proposal process takes to produce one.
