Agent Skills moves to a neutral home as Anthropic donates…

Anthropic handed its Agent Skills format to the Agentic AI Foundation on October 22, joining MCP under vendor-neutral governance. More than 70 platforms already read it — sharpening the build-vs-buy question for small-business owners.

Anthropic donated Agent Skills to the Agentic AI Foundation on October 22, 2026, handing the specification, logo, GitHub organization, and the agentskills.io domain to the Linux Foundation-affiliated body it co-founded with Block and OpenAI last December. The move follows the same trajectory Anthropic walked with the Model Context Protocol: ship it inside Claude, publish it as an open standard, then surrender governance once adoption makes neutrality strategically useful.

The timing was staged. AAIF used AGNTCon + MCPCon North America in San Jose to announce four new projects alongside Agent Skills, including Open Responses from OpenAI, OpenEnv from Hugging Face, and OpenSandbox from Alibaba. “A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10,” said Manik Surtani, the foundation’s CTO and co-founder.

The adoption numbers Anthropic attached to the handoff are striking and, per i-SCOOP, unverified: more than 70 agent platforms reading the format, roughly 1 million skills published, nearly 280 million skills installed across ChatGPT, Cursor, GitHub Copilot, VS Code, Claude Code, and the rest. Agents consume only about 100 tokens at startup, scanning each skill’s 64-character name and 1,024-character description before deciding what to load.

David Soria Parra, who co-created MCP and now serves as a Lead Maintainer of Agent Skills alongside Mahesh Murag, framed the premise plainly: “Without the right context, even the most intelligent agents fall short.” Both maintainers, along with contributors from Google and OpenAI, participate as individuals rather than on behalf of employers, under Apache 2.0 and CC BY 4.0 licensing.

The security picture is less tidy. A February 2026 Snyk analysis of 3,984 public skills across ClawHub and skills.sh found 13.4 percent carrying critical issues and 76 with confirmed malicious payloads. Cloud Security Alliance has already published supply-chain recommendations.

For a founder running a one-to-four-person company, the standardization reframes the build-vs-buy question. SKILL.md files for prospect research, personalized outreach, and content aren’t a weekend project once they’ve to be kept current, audited, and portable across seventy platforms. LemonLime sells the done-for-you side of that equation; the standard decides who else can carry the work. Standards don’t make skills good. They just make bad ones travel faster, which is the lesson MCP’s own agentic-commerce aftermath is still teaching, and the one implied by the 8-in-10 invisibility benchmark.

Agent Skills moves to a neutral home as Anthropic donates the standard to the Linux Foundation

Sources