Anthropic Donates Agent Skills Standard to New AI Foundation

Anthropic has given up control of Agent Skills, the open format it created for packaging specialized know-how into AI agents, handing the standard's name, logo, GitHub organization, and agentskills.io domain to the newly expanded Agentic AI Foundation (AAIF). The move, announced by Anthropic on October 22, 2026, follows the same playbook Anthropic used with the Model Context Protocol (MCP): build a standard inside Claude, open it up, then transfer ownership to a vendor-neutral body so no single company controls where agent interoperability goes next.

What Anthropic actually donated

The donation is narrower than it might sound. Anthropic is not changing how Agent Skills work — the format itself, a folder containing a SKILL.md file plus optional scripts, references, and assets, stays exactly as it was. What moves to the AAIF is the project's public identity and infrastructure: the Agent Skills name and logo, the GitHub organization that hosts the reference implementation, and the agentskills.io domain where the specification is published. Existing skills continue to work in Claude and in every other tool that has already implemented the format; nothing breaks for developers who published a skill before the announcement.

This is the same pattern Anthropic followed when it handed MCP, its protocol for connecting AI models to external tools and data (see our explainer on how MCP works), to the AAIF. Anthropic co-founded the foundation together with Block and OpenAI specifically to hold standards like these, so that governance doesn't sit inside the company that happened to invent the format. The Agentic AI Foundation describes its mission, on its own site, as supporting "the people and projects building agents in the open," organized around a handful of cross-company working groups covering areas like security, identity, and observability for agentic systems generally — not Agent Skills alone.

Notably, Anthropic's announcement is explicit that the donation does not hand over the specification's technical content. The AAIF's role is custodial and procedural: it owns the project's public face and provides a neutral venue for proposing changes, not a mandate to rewrite how skills work on its own authority.

What Agent Skills actually are

A skill, in Anthropic's own definition, is a folder with a SKILL.md file carrying at minimum a name and a description, plus whatever scripts, reference documents, or templates a task needs. Agents load skills through what Anthropic calls progressive disclosure: at startup an agent sees only each skill's name and short description; when a task matches one, the agent reads the full instructions into context; only then, if the skill calls for it, does the agent execute bundled code or pull in reference files. The design goal is to let an agent keep hundreds of skills "on the shelf" without burning context window on all of them at once.

Anthropic built the format while developing Claude Code, reasoning that general-purpose language models handle broad knowledge work well but lack the procedural, company-specific knowledge that real business tasks require — how a particular legal team wants contracts reviewed, how a finance team formats a specific report, which steps a support workflow always follows. Skills launched inside Claude Code in October 2025 and were published as an open, implementation-agnostic standard two months later, in December 2025, in a post Anthropic titled "Equipping agents for the real world with Agent Skills". The full technical specification — covering the exact required and optional fields in a SKILL.md file, naming conventions, and how an agent should resolve bundled scripts — is published on agentskills.io's specification page, free for any developer or company to implement without a license or sign-off from Anthropic.

That openness is also what let the format spread so quickly. Because a skill is just a folder of plain-text instructions and files rather than code tied to Anthropic's own API, any agent framework capable of reading files and following instructions can implement support for it, which is a large part of why rival products picked it up within weeks of the December 2025 publication rather than the months or years a tightly coupled, proprietary format might have taken.

Why hand it to a neutral foundation

Anthropic frames the donation as a commitment to keeping Agent Skills "open, community-driven, and vendor-neutral" now that it has become a cross-industry format rather than a Claude-only feature. The logic mirrors standard-setting in other parts of the software industry: once dozens of competing companies build products around a specification, leaving that specification's name, trademark, and governing repository under the originating vendor's control creates an obvious conflict of interest, and competitors have less incentive to invest in a format one rival can unilaterally change. Moving the project to a Linux Foundation body answers that concern in the same way it did for MCP, and for other infrastructure standards the Linux Foundation has hosted over the years.

Practically, the AAIF's role for Agent Skills going forward covers the parts of a standard that benefit from shared stewardship: a public roadmap for versioned releases, a formal process for proposing changes to the spec, and security guidance for people writing and installing skills — an area that matters more as skills increasingly bundle executable code. Anthropic has said anyone can keep using or contributing to the standard without joining the AAIF itself; discussion happens on the project's GitHub and Discord.

Who's already using Agent Skills

Anthropic's announcement cites more than 70 agent platforms as having integrated Agent Skills, naming ChatGPT, Cursor, GitHub Copilot, and VS Code specifically. Microsoft's tools were among the earliest adopters after the December 2025 open-standard release, wiring skills support into both VS Code and GitHub Copilot. Coding agents including Cursor have also adopted the format, letting the same skill folder work across multiple products instead of being rewritten per platform — the interoperability pitch that is the whole point of publishing a format as an open standard rather than keeping it proprietary.

The list of adopters matters more for what it implies than for its raw size: Anthropic, OpenAI, and Microsoft are direct competitors in the agent and coding-assistant market, and all three now support a format that one of them originally built and controlled. That's the practical test of whether a donation like this one is more than symbolic — rival companies kept building on the format after a competitor open-sourced it, rather than forking it or ignoring it in favor of their own proprietary equivalent.

Anthropic's adoption figures — the 70-platform count, roughly 1 million published skills, and nearly 280 million installs — are the company's own numbers, stated in its donation announcement. No independent organization has published a separate count, so the figures should be read as Anthropic's self-reported usage data rather than an audited total.

From launch to open standard: a timeline

Who runs the standard now

Day-to-day maintenance passes to engineers from Anthropic, Google, OpenAI, and other companies, who contribute as individuals rather than as representatives of their employers — a structure meant to keep any one company from steering the spec through its staff. David Soria Parra and Mahesh Murag hold the Lead Maintainer roles. Beyond that, the AAIF's own site describes itself around a set of cross-company working groups (covering areas like security and privacy, identity and trust, and observability) rather than publishing a detailed membership chart or voting structure for Agent Skills specifically; where the foundation hasn't published those governance details, this piece doesn't fill them in.

Discussion and proposed changes to the specification happen in the open, on the project's GitHub organization and Discord server, rather than through a closed Anthropic process. That's a meaningful shift for anyone who has watched a vendor-controlled format drift to favor its creator's own product over time: the formal complaint channel for "this change benefits Company X more than everyone else" now runs through a foundation whose working groups are explicitly cross-company, rather than through Anthropic's own product roadmap.

Security guidance is still a work in progress

Because a skill can bundle executable scripts alongside its instructions, installing one carries some of the same risk as installing any third-party code: a malicious or careless SKILL.md package could, in principle, ask an agent to run something harmful. Anthropic's donation announcement lists security guidance for skill authors and installers as one of the specific items on the AAIF's proposed roadmap, alongside versioned releases and a formal change-proposal process — but as of the donation, that guidance has been announced as a priority, not published as a finished document. Teams installing third-party skills today are, for now, relying on the same judgment they'd apply to any other code pulled from a public repository: checking the source, reading the bundled scripts before running them, and preferring skills from known or verified publishers.

What it means for developers and users

For anyone already building or installing skills, the practical answer is: nothing changes today. A SKILL.md file written against the current specification keeps working in Claude, ChatGPT, Cursor, VS Code, and GitHub Copilot exactly as before. What the donation changes is who shapes the format's future — version changes, security requirements, and the formal process for proposing new features now run through the AAIF's open GitHub and Discord channels rather than through an Anthropic-only roadmap.

For developers choosing between agent platforms — the kind of decision we've covered in our comparison of Claude Cowork and ChatGPT's agent tools — a vendor-neutral Agent Skills standard lowers the switching cost: a skill built once for internal workflows should, in principle, travel to whichever agent product a team adopts next, rather than locking a company into a single vendor's proprietary extension format. That portability is also the reason searches for "claude agent skills" and "agent skills" have been climbing: the format started as a Claude-only feature and is now something people encounter inside other companies' tools as well.

What's next

Anthropic's announcement points to a public roadmap covering versioned releases, a formal specification-change proposal process, and security guidance for people authoring and installing skills — the last of which matters more as skills increasingly bundle executable scripts rather than just reference text. None of that roadmap has shipped yet as of the donation announcement; it's a stated intention, not a published schedule. Expect the AAIF's GitHub organization and Discord to be where the first concrete changes to the specification surface, and expect more agent platforms to add Agent Skills support as the format's governance moves further away from any single vendor.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage