Anthropic Donates Agent Skills to the Linux Foundation: 1M Skills, 280M Installs, One Open Standard

What Happened: The Donation

Anthropic announced on October 22 that it is donating Agent Skills — the open standard that gives AI agents specialized knowledge and workflows — to the Agentic AI Foundation (AAIF), the Linux Foundation-backed neutral home for agent interoperability standards that Anthropic co-founded with Block and OpenAI, and to which it previously donated the Model Context Protocol (MCP).

The handover isn't symbolic. Anthropic is transferring the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain — everything needed for the community, not one vendor, to steer the standard's future. Engineers from Anthropic, Google, and OpenAI will maintain the specification as individual contributors, with David Soria Parra and Mahesh Murag serving as Lead Maintainers.

In an AI industry better known for walled gardens, this is the second major protocol Anthropic has given away in under two years — and it lands at a moment when the "agent" layer of AI tooling is exploding with incompatible, platform-locked extensions.

What Exactly Is an Agent Skill?

An Agent Skill is deliberately simple: a folder containing a SKILL.md file that teaches an agent a specific task or domain, plus any scripts or reference material it needs. A skill might encode how to run a financial reconciliation workflow, apply a company's brand style guide, operate a CAD tool, or triage a support inbox — the procedural knowledge that generic models consistently lack.

The format was born inside Claude Code, Anthropic's command-line coding agent launched in October 2025. Claude handled general knowledge work well but stumbled on the step-by-step, business-specific procedures that real jobs require. Within weeks of launch, developers had published thousands of skills on GitHub and — without any prompting from Anthropic — wrote loaders for Codex, Cursor, and OpenCode. That organic adoption convinced Anthropic to publish Agent Skills as an open standard under its own GitHub organization and domain in December 2025, so any agent could use it.

That bet paid off. Today, more than 70 agent platforms — including ChatGPT, Cursor, GitHub Copilot, and VS Code — have integrated the standard. A skill you write for one agent now travels, largely unchanged, to most of the others.

The Scale: 1 Million Skills, 280 Million Installs

The adoption numbers Anthropic disclosed are the clearest signal yet that skills have become infrastructure rather than a Claude-specific feature:

• 1 million skills published since the standard went public in December 2025
• Nearly 280 million installs across the ecosystem
• 70+ agent platforms with native or loader-based integration
• Thousands of community skills appeared within weeks of Claude Code's launch — before the standard even existed

For comparison, that install base approaches the scale of a mid-tier app store — except the "apps" are portable instruction packages that work across competing AI products from OpenAI, Anthropic, Google, and dozens of independents.

Why Vendor Neutrality Matters for AI Tools

If Agent Skills had remained an Anthropic project, every competitor integrating it would have been depending on a rival's roadmap, branding, and goodwill. That hesitation is exactly what kills emerging standards. Moving the trademark, the GitHub org, and the domain to a Linux Foundation home removes the single-vendor risk for the 70+ platforms already shipping support — and clears the path for Google, OpenAI, and others to contribute as peers rather than adopters.

The precedent is MCP, which followed the same path: launched by Anthropic, adopted across the industry, then donated to the AAIF. That sequence turned MCP from "Anthropic's protocol" into the default way agents connect to tools and data. Agent Skills is now positioned to do the same for the layer above it — how agents acquire knowledge and procedures rather than connections.

For anyone choosing AI tools in 2026, the practical effect is simple: the skills you write and install are no longer tied to any single vendor's survival or strategy. A tool that supports Agent Skills today is a safer long-term bet than one with a proprietary-only extension format.

The Roadmap: Versioning, Proposals, Security

Anthropic has proposed the foundation start with a roadmap covering three priorities:

• Versioned releases — so platforms can pin against a stable spec instead of a moving target
• A formal proposal process — community-driven changes with transparent decision-making, following the MCP model
• Security guidance — best practices for both the people who write skills and the people who install them

The security piece deserves attention. A SKILL.md file is executable instructions delivered to an agent that may have file-system, network, and API access — which makes skill marketplaces a genuine supply-chain surface. Formal security guidance, and eventually signing or attestation conventions, are what separate a hobby format from enterprise-grade infrastructure. Anyone can join the project's Discord to weigh in on proposed specification changes, and you don't need to join the AAIF to use or contribute to the standard.

MCP and Agent Skills: Two Halves of the Agent Stack

With both standards now under one roof, the agentic stack is consolidating into two clean layers:

An agent that speaks MCP can reach your systems; an agent with the right skill knows what to do once it gets there. Together they cover most of what "agentic AI" promises — which is why their shared neutral home matters.

What Tool Builders and Users Should Do Now

If you build or choose AI tools, a few concrete takeaways from this week's announcement:

• Tool makers: if your product has an agent or assistant, supporting the Agent Skills format (a SKILL.md loader is a weekend of work) instantly plugs you into a million-skill ecosystem — and it's now vendor-neutral, so there's no lock-in objection left.
• Teams using agents: encode your internal workflows as skills once, and they'll port across Claude, ChatGPT, Copilot, Cursor, and more — instead of rewriting prompts per platform.
• Skill authors: watch the AAIF's versioning and security roadmap; pin your skills to versioned releases once they exist, and follow the upcoming security guidance before distributing anything that runs scripts.
• Everyone else: when comparing AI tools, treat open-standard support (MCP, Agent Skills) as a durability signal — proprietary extension formats are a switching-cost trap.

The specification lives at agentskills.io, development happens on GitHub under the agentskills organization, and the project Discord is open for feedback on proposed changes.

Frequently Asked Questions