On October 22, 2026, Anthropic handed the Agent Skills standard (the format behind every SKILL.md) to the Agentic AI Foundation, the Linux Foundation project that already hosts MCP. If you install or write skills, nothing you do today has to change. But the move puts a spotlight on a rule most skills ignore: the standard allows six frontmatter fields, and anything else breaks a skill outside Claude Code. We checked every skill in our catalog against it.
Key Takeaways
- Anthropic donated Agent Skills to the Agentic AI Foundation (AAIF) on October 22, 2026. The name and logo, the GitHub organization and the agentskills.io domain moved. The AAIF sits under the Linux Foundation and also hosts MCP and AGENTS.md.
- The format didn't change. The specification was last edited on August 4, 2026. A skill is still a folder with a
SKILL.mdfile. - The roadmap is a proposal. Anthropic proposed versioned releases, a formal proposal process and security guidance as the first work. None of it is published yet.
- The standard allows six frontmatter fields. Claude Code accepts 20. Uploading a skill to claude.ai or the Skills API with any of the other 14 fails with a hard error.
- 12 of the 158 skills in our catalog with frontmatter use a field outside the spec, including shadcn/ui, Impeccable and all three of Expo's skills.
What Happened on October 22
Anthropic announced it was donating Agent Skills to the Agentic AI Foundation. The donation covers the Agent Skills name and logo, the agentskills GitHub organization (where the specification lives) and the agentskills.io domain.
The standard is maintained by engineers from Anthropic, Google, OpenAI and other companies, acting as individuals. David Soria Parra and Mahesh Murag are the Lead Maintainers. Soria Parra is also co-creator and lead maintainer of MCP. You don't have to join the foundation to use the standard or contribute to it.
The same day, the Linux Foundation announced four new AAIF projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba.
Anthropic's post puts adoption at more than 70 agent platforms, among them ChatGPT, Cursor, GitHub Copilot and VS Code. It also counts 1 million published skills and nearly 280 million installs. Those are Anthropic's figures; we haven't been able to check them.
What Is the Agentic AI Foundation?
The AAIF is a Linux Foundation project that holds open standards and infrastructure for AI agents, so that no single company controls them. Anthropic co-founded it with Block and OpenAI, and it launched in December 2025 with three projects: MCP, Block's goose and OpenAI's AGENTS.md. With the four new projects it now hosts ten.
For skill users, the practical point is that the three pieces most agents now share live under one roof: MCP for tools, AGENTS.md for repo instructions, and Agent Skills for packaged know-how. If you're unsure which one you need, our comparison of skills and MCP servers covers it.
What Didn't Change: The Format
The specification is the same document it was before the announcement. Its last edit landed on August 4, 2026, and the repository's last commit was on August 9. A skill is still a directory with a SKILL.md file: YAML frontmatter, then Markdown instructions, with optional scripts/, references/ and assets/ folders beside it.
The frontmatter allows six fields:
| Field | Required | What the spec says |
|---|---|---|
name | Yes | Up to 64 characters: lowercase letters, numbers and single hyphens. Must match the parent directory name. |
description | Yes | Up to 1,024 characters. What the skill does and when to use it. |
license | No | A license name, or a reference to a bundled license file. |
compatibility | No | Up to 500 characters on environment requirements. |
metadata | No | Any key-value pairs, as strings. This is where a version number belongs. |
allowed-tools | No | Tools the skill may use without asking. Marked experimental. |
Nothing you've installed stops working, and every install command on this site still works. If you're new to the format, our guide to what skills are starts from the beginning.
What's Proposed: Versions, Proposals, Security Guidance
Anthropic proposed three things as the first work under the foundation. Each is a proposal; none has been published.
- Versioned releases. Today the spec has no version number, so “follows the standard” can't name which standard. With versions, a client could state which spec version it supports, and an author could target one.
- A formal proposal process. Changes to the format would go through a written proposal and review, as MCP's spec changes already do.
- Security guidance for people who write and install skills. The spec says almost nothing about security today. Skills can bundle scripts and, in Claude Code, pre-approve tools and register hooks. Guidance on what to check before installing one is overdue.
We'll update this page when any of the three ships.
The Standard vs. Claude Code's Extensions
This is the part of the news that affects skills today. Claude Code's skills documentation says its skills follow the Agent Skills standard and that Claude Code “extends the standard.” It accepts 20 frontmatter fields. The other 14 are Claude Code's own: among them hooks, disable-model-invocation, user-invocable, argument-hint, model, effort and context.
Those fields work in Claude Code, including in plugin skills. Outside it, they don't. A claude.ai skill upload, the Skills API and packaging with package_skill.py accept only the six spec fields. The docs say any other field fails “with a hard error instead of ignoring the field”:
Unexpected key(s) in SKILL.md frontmatter: argument-hint. Allowed properties are: allowed-tools, compatibility, description, license, metadata, name
That includes uploading a personal skill to your claude.ai account so you can use it in Cowork, cloud sessions or scheduled tasks. Body features that only Claude Code has, such as !`command` dynamic context, don't work in claude.ai chat or through the API either.
12 Skills in Our Catalog Use Fields Outside the Spec
On October 26, 2026, our catalog listed 193 skills, and we carried the SKILL.md text for 160 of them. 158 of those have frontmatter. We listed every top-level frontmatter key in each one and compared the keys with the spec's six. 12 skills use at least one key the spec doesn't allow:
| Skill | Field(s) outside the spec |
|---|---|
| Impeccable | version, user-invocable, argument-hint |
| shadcn/ui | user-invocable |
| Interface Review | disable-model-invocation |
| Property-Based Testing | effort |
| Resend | inputs, references |
| Expo App Design, Expo Deployment, Upgrading Expo | version |
| Remotion Best Practices | version |
| iOS Simulator | version |
| Varlock | version |
| ClawSec | version, homepage, clawdis |
That isn't a verdict on these skills. All of them load in Claude Code, and the fields do real work there. But if you try to use one of them in claude.ai or through the Skills API, expect the unexpected-key error until its author changes it. The most common culprit, a top-level version, has a simple fix: move it under metadata, which is where the spec's own example puts it.
The spec also requires a skill's name to match the folder it sits in. Nine skills in our catalog that live in a subfolder of their repo don't, among them Taste Skill (design-taste-frontend in skills/taste-skill/) and four of Vercel's React skills, whose names carry a vercel- prefix their folders don't. We know of no client that rejects that today. The spec's reference validator does flag it.
What to Do Now
If you install skills
- Nothing. Installed skills keep working, and install commands haven't changed.
- If you use skills in claude.ai or through the Skills API, check the frontmatter of a skill before you upload it. The table above lists the ones in our catalog that will fail.
If you write skills
- Keep to the six fields if the skill should work beyond Claude Code. Put a version and anything else of your own under
metadata. - If you need Claude Code's fields, such as
hooksordisable-model-invocation, say in your README that the skill is for Claude Code. - Validate before you publish. The spec repository ships a reference library,
skills-ref, whoseskills-ref validate path/to/skillchecks the name rules, the folder match and the allowed fields. Its README says it is for demonstration, not production use, but it is the closest thing to a conformance test the standard has. - Watch the proposal process. Once versioned releases arrive, that is where a new field or a changed rule will show up first.
New to writing skills? Our step-by-step guide builds one from scratch.
Frequently Asked Questions
How We Verified This
Checked on October 26, 2026. The donation, the maintainers and the roadmap come from Anthropic's announcement of October 22. The foundation's history and the four new projects come from the Linux Foundation's press release the same day. The spec's fields, its last edit date and the validator's checks come from the agentskills/agentskills repository: docs/specification.mdx and skills-ref/src/skills_ref/validator.py. The 20 fields Claude Code accepts, and the upload error, are quoted from Claude Code's skills documentation.
The catalog counts come from the SKILL.md text we fetched from each skill's upstream repository on October 26, 2026. We counted top-level frontmatter keys only. Skills change often; if one in the table has since been fixed, tell us.
Now It's Your Turn
The move to the Linux Foundation changes who decides what a skill is, not what a skill is. The format is the same file it was in August. What's worth doing now is small: keep your frontmatter to the spec's six fields if your skill should work beyond Claude Code, and run the spec's validator before you publish. Versioned releases will make that check matter more, not less.
Browse All Skills