On October 22, Anthropic announced that it is donating Agent Skills to the Agentic AI Foundation (AAIF), the Linux Foundation vehicle that already hosts MCP. It is the second time Anthropic has handed a core agent primitive to neutral governance, and it matters more than it looks if you write specs for a living.
Caveat up front: I am working from Anthropic’s announcement and the linked project pages. I haven’t independently audited any of the adoption numbers below.
What actually moved#
According to the post, the donation covers four things:
- The Agent Skills name and logo
- The GitHub organization
- The agentskills.io domain
- Stewardship of the standard itself
The format didn’t change. A skill is still a folder with a SKILL.md file, plus whatever scripts and reference material the task needs. Anthropic says it came out of Claude Code, where general-purpose models kept needing procedural knowledge: how your company does a release, how your repo wants migrations written, what “done” means for a data pipeline.
Anthropic’s stated commitment is to keep the standard “open, community-driven, and vendor-neutral,” and you don’t need to join AAIF to use or contribute to it.
The numbers (vendor-reported)#
Anthropic’s post makes these claims:
- Skills launched in October 2025. Within weeks, developers had published thousands on GitHub and written loaders for Codex, Cursor and OpenCode.
- The standard was published openly in December 2025.
- 70+ agent platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code.
- 1 million skills published, with nearly 280 million installs.
Treat “installs” with the usual suspicion: it’s a download counter, not a usage counter. The directional claim is still credible, because I can see the loaders in the competing tools myself. Cursor and Copilot reading SKILL.md folders is the strongest evidence that the format escaped its origin.
Governance: who’s in the room#
Per the announcement, maintainers from Anthropic, Google, OpenAI and others maintain the standard as individuals. David Soria Parra and Mahesh Murag are lead maintainers. AAIF was co-founded by Anthropic, Block and OpenAI, and Anthropic previously donated MCP to it.
The proposed roadmap has three items:
- Versioned releases. Today a skill is effectively “whatever the loader understands this week.”
- A formal proposal process, so changes to the format go through review rather than a blog post.
- Security guidance for people who write and install skills.
The post gives no further technical details on the spec, so I won’t speculate about field changes.
Why SDD people should care#
Spec-Driven Development has a distribution problem. You can write a superb spec, but if it only works in one tool’s proprietary rules format, you’ve built a lock-in, not a methodology. A skill is the closest thing we have to a portable unit of spec: a human-readable, version-controllable folder that tells an agent how a task is done here.
Putting it under the Linux Foundation changes the risk calculation for adopters:
- Teams can invest in a shared skills library without betting on one vendor’s roadmap. I covered the mechanics of that in scaling Claude Code skills across an engineering org.
- Tool vendors lose the excuse for a bespoke format. If Copilot, Cursor and Claude Code all read the same folder, differentiation moves to the agent loop, where it belongs.
- Security teams get a venue for the supply-chain problem. Skills are executable instructions plus scripts, installed from the internet. The roadmap’s security-guidance item is overdue, and I’ve already written about skill scanners and the supply-chain risk.
If you want the broader picture of where skills sit relative to MCP and plugins, see skills, plugins and MCP: the three extension layers.
My take: good move, with an honest asterisk#
I’m pro-Anthropic, and this is a genuinely good move. Giving away a standard that competitors already use costs Anthropic little and cements the ecosystem around the format it invented. It also mirrors the MCP playbook that made MCP the default: ship something useful, let everyone adopt it, then neutralize the “it’s just Anthropic’s thing” objection.
The asterisk: neutral governance doesn’t fix the quality problem. A million published skills is a million unreviewed instruction sets, and “install” is a trust decision most people make in two seconds. Versioning and proposal processes are necessary; they aren’t sufficient. Until there’s signing, provenance and a sane review story, treat third-party skills like third-party npm packages, because that’s what they are.
Contrast this with the IDE-centric camp, whose rules formats remain tool-specific and whose agents still assume a human watching the editor. A portable, vendor-neutral skill folder is exactly what an autonomous agent needs: knowledge that travels with the repo, not with the IDE.
What to do this week#
- Audit your skills folder. Anything you’ve installed from a third party should be read like code, because it is.
- Write your team’s top three procedures as skills: release, migration, incident triage. Keep each under a page.
- Pin and review. Check skills into the repo and review changes in PRs, rather than pulling them ad hoc.
- Watch the AAIF roadmap for the versioning and security-guidance deliverables; those are what will make this safe to depend on.
Sources#
- Anthropic, “Donating Agent Skills to the Agentic AI Foundation” (Oct 22, 2026)
- Anthropic, “Donating the Model Context Protocol and establishing the Agentic AI Foundation”
- agentskills.io and the specification discussion repo
