Skip to content
audiotale.net
Sign in
NewsNews

Anthropic donates Agent Skills (SKILL.md) to the Agentic AI Foundation

The open skills format used by 70+ agent platforms now sits under the Linux Foundation's AAIF; what changes today, what is only promised, and what we couldn't verify.

audiotale.net team

· 7 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. The numbers
  5. How it works
  6. What it means
  7. What we don't know yet
  8. What to watch next
  9. Sources

Anthropic is donating Agent Skills, the folder-of-instructions format that coding agents use to load specialised know-how on demand, to the Agentic AI Foundation (AAIF), a body hosted by the Linux Foundation. The announcement came on 22 October 2026. The standard's name, logo, GitHub organisation and the agentskills.io domain move to the foundation, and engineers from Anthropic, Google, OpenAI and other companies will maintain it as individuals. For developers the format itself does not change today; what changes is who controls it.

Key points

  • Anthropic announced on 22 October 2026 that it is donating the Agent Skills standard, name, logo, GitHub organisation and agentskills.io domain to the Agentic AI Foundation, hosted by the Linux Foundation.
  • David Soria Parra and Mahesh Murag are the Lead Maintainers; engineers from Anthropic, Google, OpenAI and other companies will maintain the standard as individuals.
  • Anthropic says more than 70 agent platforms have integrated the standard, about 1 million skills have been published and nearly 280 million installed; these are Anthropic's figures.
  • The proposed roadmap covers versioned releases, a formal proposal process and security guidance for people who write and install skills; none of that has shipped yet.
  • Nothing in the format changes: you can keep using or contribute to Agent Skills without joining the foundation.

What happened

On 22 October 2026 Anthropic published "Donating Agent Skills to the Agentic AI Foundation". The same day, at AGNTCon + MCPCon North America in San Jose, California, the AAIF announced four new projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba.

According to Anthropic, the donation covers the Agent Skills standard itself, its name and logo, its GitHub organisation and the agentskills.io domain. Anthropic says it is "committed to keeping Agent Skills open, community-driven, and vendor-neutral." The AAIF quotes Soria Parra: "Today, we're donating Agent Skills to the AAIF to ensure that anyone building with skills can shape what's next."

Anthropic describes governance this way: engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals, with Soria Parra and Murag as Lead Maintainers. The first priorities, it says, are a roadmap with versioned releases, a formal proposal process for changes, and security guidance for skill authors and installers. Feedback on proposed specification changes goes through the project's Discord; the specification is at agentskills.io and discussion happens on GitHub.

Anthropic's announcement does not state licence terms, and the AAIF's announcement does not either. A developer-focused write-up from Developers Digest likewise says the spec's licence is not stated in the coverage it read and tells readers to check the repository. We did not verify the licence, so check it before depending on it.

Background

A skill, in Anthropic's description, is a folder containing a SKILL.md file plus any scripts or reference material it needs. The agent reads a short description of each installed skill and loads the full instructions only when a task calls for them, which keeps the context window small. Anthropic says it developed the idea while building Claude Code and launched skills in October 2025. It published the format as an open standard in December 2025, under its own GitHub organisation and domain. Anthropic says developers wrote loaders for Codex, Cursor and OpenCode within weeks.

The foundation is a few months younger than the standard. On 9 December 2025 the Linux Foundation announced the Agentic AI Foundation, with Anthropic contributing the Model Context Protocol (MCP), OpenAI contributing the AGENTS.md specification and Block contributing the open-source Goose framework, according to SiliconANGLE's report that day. Platinum members listed there are Anthropic, Block, OpenAI, Amazon Web Services, Bloomberg, Cloudflare, Google and Microsoft. Anthropic says it co-founded the AAIF with Block and OpenAI and previously donated MCP to it, so Agent Skills follows the same path as MCP.

Developer Simon Willison had predicted this. He wrote in December 2025 that he "wouldn't be surprised to see this end up in the AAIF", and noted the donation in a short post on 22 October.

Developers Digest reports that the AAIF now hosts 10 projects, up from three at launch. That count comes from the write-up; we did not find it in the AAIF's own announcement.

The numbers

Anthropic's adoption figures: more than 70 agent platforms have integrated the standard since December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code; about one million skills have been published; and nearly 280 million installs. Two independent write-ups, Developers Digest and Mixed, both flag these as self-reported by Anthropic. Mixed adds that the install figure is not an independent count of distinct users, and we agree: "installs" could include repeat installs, and nobody outside Anthropic has audited it.

There is one independent number about safety. Developers Digest cites a scan by "The Agent Forge" of 9,930 public SKILL.md files from 277 repositories. By its method the scan found no malware but flagged 30 files (0.33%). Fourteen files piped a download into a shell, and eight read a whole key file to get one API key. Of the 30 flags, the write-up says 26 were real patterns with no sign of malice. We have not seen the scanner's own report, so this is a second-hand figure, and the scanner's makers admit it can be evaded, for example by splitting a payload across files. A low flag rate on public repositories says little about private or brand-new skills.

SKILL.md files flagged in a scan of 9,930 public files (30 flagged, 0.33%)files
SKILL.md files flagged in a scan of 9,930 public files (30 flagged, 0.33%) (files)
files
Download piped into a shell14
Whole key file read for one API key8
Other flagged files (remainder)8

Source: The Agent Forge scan, as reported by Developers Digest

How it works

A skill is deliberately simple. The directory holds a SKILL.md with a name and description at the top, followed by instructions in Markdown, and optionally scripts and reference files. Developers Digest summarises the specification rules: name is 1 to 64 characters of lowercase letters, numbers and hyphens, matching the parent directory; description is required and up to 1,024 characters; optional fields include license, compatibility, metadata and an experimental allowed-tools. That article's author notes that a skills-ref validate command appears in the spec but was not run, so confirm it in the README before adding it to continuous integration.

Because a skill can include scripts, installing one is closer to installing a package than to adding a document. That is why the planned security guidance matters. Mixed notes that individual skills can still depend on particular tools and permissions, so a shared format does not by itself make them portable or safe.

What it means

For developers who write skills: the practical effect is stability of governance. A format controlled by one vendor can be changed by that vendor; a format under a foundation, with maintainers drawn from several companies and a formal proposal process, is easier to bet a team's workflows on. Nothing needs migrating today.

For tool makers: with Google and OpenAI engineers among the maintainers, adding or keeping skills support is less of a competitive concession than it was. Anthropic names ChatGPT, Cursor, GitHub Copilot and VS Code among the platforms that already support it.

For security teams: skills are executable-adjacent content distributed at large scale. Versioned releases and security guidance are promised, not delivered. Until then, treat third-party skills like third-party code: read them, pin versions where you can, and be wary of ones that pipe downloads into a shell.

For Anthropic: it gives up control of a format that others already adopted, which mirrors what it did with MCP. Anthropic frames it as keeping the standard open and vendor-neutral.

What we don't know yet

  • The licence of the specification. Neither announcement states it.
  • How decisions will be made beyond the proposed roadmap: the voting or approval process, and how individual maintainers are added or removed.
  • Whether the one-million and 280-million figures count distinct skills and distinct installs, since they are Anthropic's own.
  • When versioned releases and the security guidance will arrive.
  • How many of the 70-plus platforms support the full format rather than a subset.

What to watch next

Watch the agentskills.io specification and its GitHub organisation for the first versioned release, the proposal process and the security guidance. Watch whether other major agent platforms that don't yet list support add it now that governance is neutral. And watch for independent audits of public skills: the one scan we found covered 9,930 files from 277 repositories, a small slice of a catalogue Anthropic puts at a million.

For background on how software updaters are made safe, see how secure software updates work.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next