Anthropic has given the Agent Skills standard to the Agentic AI Foundation, the Linux Foundation body that already hosts MCP. Nothing in a working SKILL.md breaks today. What changes is who decides the next version, so teams that share skills across agents should check them against the published spec and plan for versioned releases.
What was announced
The handover was announced by Anthropic on 22 October 2026 in Donating Agent Skills to the Agentic AI Foundation. Anthropic first published Agent Skills as an open standard in December 2025, and the announcement says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code have integrated it since then.
Four things move to the foundation: the Agent Skills name and logo, the agentskills GitHub organisation, and the agentskills.io domain where the specification is published. The Agentic AI Foundation (AAIF) was co-founded by Anthropic, Block and OpenAI, and it is also the home of the Model Context Protocol.
What actually changed
The format itself did not change with the announcement. The changes are about governance:
- Maintainers act as individuals. Engineers from Anthropic, Google, OpenAI and other companies maintain the standard in a personal capacity, with David Soria Parra and Mahesh Murag as Lead Maintainers.
- No membership is needed. Anthropic states that anyone can use the standard or contribute to it without joining AAIF.
- A roadmap is proposed, not shipped. The announcement lists versioned releases, a formal proposal process, and security guidance for people who write and install skills as planned work. None of these exists yet.
The specification in the agentskills repository is the reference to build against. Its code is licensed under Apache 2.0 and its documentation under CC-BY-4.0, according to the repository README. The repository currently has no tagged releases, so there is no version number to pin.
What the specification requires today
A skill is a folder with a SKILL.md file, plus optional scripts/, references/ and assets/ directories. The specification sets these rules for the YAML frontmatter:
nameis required: 1 to 64 characters, lowercase letters, digits and single hyphens, not starting or ending with a hyphen, and it must match the folder name.descriptionis required: 1 to 1,024 characters, saying what the skill does and when to use it.license,compatibility(up to 500 characters) andmetadata(string keys to string values) are optional.allowed-toolsis optional and marked experimental, and the specification says support for it varies between agents.
The specification also describes progressive disclosure: agents load only name and description for every skill at startup, load the full SKILL.md body when the skill is activated, and read other files only when a task needs them. It recommends keeping SKILL.md under 500 lines and file references one level deep. The repository includes a reference validator, skills-ref validate ./my-skill, that checks frontmatter and naming.
What it means for teams building AI agents
For most teams this is a good moment for a short audit rather than any migration. CodeDTX's advice:
Validate every skill you share across tools. Agents written before the specification settled have accepted looser files. Run skills-ref validate in CI on your skills repository, so a folder name that doesn't match name, an uppercase letter or a double hyphen fails the build before it reaches a client that is stricter than the one you tested with.
Keep client-specific behaviour in metadata, not in new top-level fields. A foundation with a proposal process means new top-level fields will be agreed in public. If you invent your own today, a future version may give the same name a different meaning. The specification already provides metadata for this, and suggests key names unique enough to avoid clashes.
Don't rely on allowed-tools for security. It is experimental and not every agent honours it. Permissions for what a skill can touch belong in the agent's own tool configuration and in the systems it calls. Our post on governing agent skills as a supply chain covers how to separate a skill's instructions from its authority.
Record a version in each skill yourself. Until the standard has versioned releases, nothing in a SKILL.md says which edition of the format it was written for. A metadata entry such as spec-checked: 2026-10 plus your own skill version gives you a trail when the first numbered release lands.
Watch the proposal process if skills are part of your product. Teams that ship skills to customers, or build an agent that loads them, now have a public route to raise compatibility needs. Anthropic points to the project's Discord for feedback on proposed changes.
When not to change anything
If your skills live inside one product, such as Claude Code alone, and you have no plan to share them, there is nothing to do this week. The same files load the same way. Revisit when the foundation publishes its first versioned release or its security guidance.
How this relates to MCP
Skills and MCP now sit under the same foundation but do different jobs: a skill tells an agent how to do a task, while an MCP server gives it tools and data to do it with. If you are deciding which one a capability needs, our guide on when an agent needs MCP or A2A covers the integration side.
Frequently asked questions
Does the move to the Agentic AI Foundation break existing SKILL.md files?
No. The announcement transfers the name, the GitHub organisation and the agentskills.io domain, not the format. A skill that follows the current specification, with a valid name matching its folder and a description, loads as before. Changes to the format will come later through the proposed versioned releases and proposal process, which do not exist yet.
Do we need to join AAIF to use or contribute to Agent Skills?
No. Anthropic states that anyone can use the standard or contribute to it without joining the Agentic AI Foundation. Feedback on proposed specification changes goes through the project's Discord, and contributions go through the agentskills GitHub repository. Joining the foundation matters only if your organisation wants a formal role in how the foundation itself is run.
Which Agent Skills version should we pin to?
There is nothing to pin yet. The agentskills repository has no tagged releases, and versioned releases are listed as proposed work. Until they arrive, check your skills against the current specification with the skills-ref validator and record the date you checked in each skill's metadata, so you can tell which files need a review when the first numbered release is published.
Can we rely on allowed-tools to restrict what a skill does?
Not on its own. The specification marks allowed-tools as experimental and says support varies between agent implementations, so one client may honour it while another ignores it. Treat it as a hint. Enforce real limits in the agent's tool permissions and in the systems the agent calls, where a misbehaving or tampered skill cannot change them.
What is the difference between Agent Skills and MCP?
A skill is a folder of instructions, scripts and reference files that tells an agent how to carry out a task. An MCP server exposes tools and data that an agent can call. Both are now hosted by the Agentic AI Foundation, and they work together: a skill can describe a workflow that uses tools an MCP server provides.
