Anthropic donated Agent Skills, the folder format that teaches an AI agent a specific task, to the Agentic AI Foundation on October 22, 2026. The transfer covers the standard’s name and logo, its GitHub organization and the agentskills.io domain, so the format now belongs to a foundation rather than to the company that also sells the models it runs on.
What was actually released
A skill is a folder holding a SKILL.md file plus whatever scripts and reference material the task needs. Anthropic says it built the format while working on Claude Code, which handled general knowledge work but lacked the procedural knowledge that real business tasks require. Developers published thousands of skills on GitHub within weeks of the October 2025 launch and wrote loaders for Codex, Cursor and OpenCode, and the format was published as an open standard in December 2025.
What moved this week is that standard, not a product release. Anthropic handed the name and logo, the GitHub organization and the agentskills.io domain to the AAIF, a body under the Linux Foundation that Anthropic co-founded with Block and OpenAI, and to which it had already given the Model Context Protocol. Anyone can keep using the standard or contributing to it without joining the foundation.
| Asset | Where it goes | What it changes for a developer |
|---|---|---|
| Standard name and logo | Agentic AI Foundation | nothing inside a skill file |
| GitHub organization | Agentic AI Foundation | same repositories, new maintainers |
| agentskills.io domain | Agentic AI Foundation | specification and discussion move there |
| Specification text | maintainers from Anthropic, Google and OpenAI, as individuals | the contribution process is proposed, not yet in force |
The assets Anthropic lists as donated. The fourth row decides how the standard evolves.
Governance is the second half of the announcement. Engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals, with David Soria Parra and Mahesh Murag named as lead maintainers. Anthropic says proposed specification changes go to a Discord discussion first.
What the claim rests on
The adoption figures are Anthropic’s own, and none of them comes with a method. The post says more than 70 agent platforms have integrated Agent Skills since December 2025. Four are named, ChatGPT, Cursor, GitHub Copilot and VS Code, and the rest are not, so the count cannot be checked platform by platform. The post also does not say whether an integration means native support or a community loader, which are different commitments.
The install figure has the same shape. Anthropic counts 1 million skills published and nearly 280 million installed, with no date range, no definition of an install, and no count of skills that still work. The number a buyer would want, how many of those skills run on more than one platform, is not published.
What can be checked on the day of the transfer is the format itself. A skill is a text file that describes a task and the material it needs, which is why the standard is portable in principle: a folder a runtime can read does not need the vendor’s permission to exist. The mechanisms that make such a format safe to depend on are exactly what the announcement leaves open.
What is not published
No date for the first versioned release under the foundation, no process for deprecating an old version, no security guidance for installing a third-party skill, and no audit of the million skills already published. The maintainers are named as individuals, so there is also no company support contract behind the standard.
“We’ve proposed starting with a roadmap covering versioned releases, a formal proposal process, and security guidance for people who write and install skills.”
Read that sentence twice. The transfer of ownership is done, the rules of the road are not.
What actually changed on October 22
- The name, the logo, the GitHub organization and the domain belong to the Agentic AI Foundation, not to Anthropic.
- The specification is maintained by named engineers from several companies, acting as individuals.
- Nothing in the standard’s content changed, and existing skills keep working without a migration step.
On this page
What it changes for teams running agents
Three consequences, in the order they reach a budget.
The first is procurement. A file format owned by the company that sells the models can change with a product decision. Foundation ownership removes that specific risk, and the precedent is already in the same family: the Model Context Protocol went to the same foundation, and MCP is now the layer of an agent stack that nobody renegotiates with a single vendor.
The second is portability, which ownership does not buy. A skill that calls a model-specific tool, or that assumes a particular runtime’s file access, will not run elsewhere just because the format is neutral. The only test that settles it is mechanical: take a skill folder your team wrote, run it on a second runtime, and see what breaks. We saw the same pattern inside one vendor when Anthropic folded subagents into a single agent loop: the packaging gets simpler and the runtime assumptions get harder to see.
The third is security, and it is the one with a date attached. The post describes a skill as a folder that carries any scripts or reference material it needs, and the third item on the roadmap is security guidance for people who install skills. Until that exists, installing a skill written by a stranger means running their code with your agent’s permissions, inside whatever repository or shell that agent can reach. Packaging know-how into installable units only helps if the units are reviewed like dependencies, which is the point we made when Microsoft put 20 legal skills inside Copilot.
Versions are the loose end. There is no versioned release of the specification yet, so a workflow that depends on a detail of the format should pin the commit it was built against, not the domain.
The Neural Minute take
The governance story is the easy one to write, and it is not the story a buyer needs.
The move is real and it is the right direction: a standard that several companies maintain is worth more than a standard that one company can revise. Our reservation is about what the announcement uses to sell it. Seventy platforms is a claim without a list, nearly 280 million installs is a claim without a definition, and both will be repeated in every headline about the move. The file format itself checks out: small, readable, vendor-agnostic.
For anyone about to build on skills: treat the format as stable and the tooling as unfinished. Keep your skills in your own repository with a review step, because a skill can ship scripts and nobody vets them for you today. Record the commit hash of the specification you targeted, because versioned releases are promised and not published. Test one real skill on a second runtime before you standardize a workflow on the format.
The part of the announcement worth watching is the proposal process, not the donation. Whoever controls it decides what a skill is allowed to do, and whether a neutral foundation ends up as a vendor roadmap with a new logo.
Sources
- Donating Agent Skills to the Agentic AI Foundation, anthropic.com, published October 22, 2026: the announcement
- Introducing the Model Context Protocol, anthropic.com, the earlier donation to the same foundation: the MCP announcement
- The figures used here (more than 70 platforms, 1 million skills published, nearly 280 million installs) are Anthropic’s own counts, published in the post above. No methodology, date range or platform list is given with them.