Anthropic's Second Donation in Under a Year
In December 2025, Anthropic donated MCP, the Model Context Protocol, to the Agentic AI Foundation (AAIF) under the Linux Foundation. Less than a year later, on October 22, 2026, it handed over Agent Skills as well — together with the standard's name and logo, its GitHub organization, and the agentskills.io domain. Two of the company's most important open standards, donated to the same neutral foundation within twelve months. In an industry where vendors habitually keep their protocols close to the chest, this is a move worth examining carefully.
First, a date correction. Posts circulating on social media placed the announcement on October 24 or 25. That is wrong. Anthropic's own newsroom page lists the announcement under the Product category with a date of October 22, 2026, and the Linux Foundation press release carries a SAN JOSE dateline of Oct. 22, 2026. Both official sources agree. The announcement was made on stage at AGNTCon + MCPCon North America, held October 22–23 at the McEnery Convention Center in San Jose — meaning the ownership changes of both MCP and Agent Skills were announced from the same podium.
Four Projects, One Day: AAIF Grows From 3 to 10
This was not a one-company show. On the same day, four vendors each donated a project: Anthropic donated Agent Skills, OpenAI donated Open Responses, Hugging Face donated OpenEnv, and Alibaba donated OpenSandbox. Manik Surtani, CTO of the AAIF, provided the tersest summary in the press release: "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10."
The four new projects divide the territory neatly. Agent Skills is the open standard for packaging specialized knowledge and workflows for AI agents. Open Responses is a multi-vendor, interoperable LLM interface specification that grew out of OpenAI's Responses API, unifying schema, streaming events, and tool-calling semantics. OpenEnv is an open framework and community hub for reinforcement learning environments, exposing agent interfaces, MCP, and the Gym training protocol through containerized services. OpenSandbox is a general-purpose sandbox platform with multilingual SDKs, a unified sandbox API, and Docker/Kubernetes runtimes for the safe execution of coding agents, GUI agents, evaluation, and training. Added to the six projects already under management — MCP, Agent2Agent (A2A), Agent Router, agentgateway, AGENTS.md, and goose — the foundation's portfolio now covers, for the first time, the protocol layer, the interface layer, the training-environment layer, and the execution layer of the agent stack.
The foundation's background matters too. The AAIF is a directed fund under the Linux Foundation, co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg. It launched with three projects — MCP from Anthropic, and, according to consistent reporting from multiple media outlets, AGENTS.md from OpenAI and goose from Block. The Linux Foundation press release confirms only the count, not the names, so we treat that list as a secondary source.
The Three Numbers: All Self-Reported
The most eye-catching part of the announcement is a trio of numbers: more than 70 agent platforms have integrated Agent Skills, about 1 million skills have been published, and nearly 280 million have been installed. All three figures come straight from Anthropic's announcement — and before quoting them, the provenance needs to be stated plainly: these are Anthropic's self-reported numbers. No independent audit or third-party statistics accompany them, and media coverage across the board simply repeats the official figures. No contradicting independent count has surfaced.
Taking them one at a time. The 70-plus platform figure has a precise official wording: "more than 70 agent platforms" have integrated Agent Skills "since we published the standard in December 2025." The time window is explicit, and the company names ChatGPT, Cursor, GitHub Copilot, and VS Code. The window itself is a form of restraint — the count starts not from the October 2025 feature launch but from the December standard publication. The 1 million published skills and nearly 280 million installed figures come with only a "so far," with no more precise cutoff given, and the counting basis for installs — per platform or per user — is unspecified. Throughout this article we therefore attribute the figures as "according to Anthropic": the numbers stand as quoted, and their weight is for you to judge.
Are self-reported numbers worthless, then? No. For ecosystem-level metrics like these, external audits are rare and vendor self-reporting is the industry norm; what you can do is probe internal consistency. The 70-plus figure names competitor platforms, which means it invites scrutiny from the very companies it counts. The window starts from the standard's publication rather than the earlier feature launch, which means Anthropic did not pad the number with its own private ecosystem. None of this proves the number is accurate, but it shows the reporter tried to make it withstand questioning — considerably more credible than a bare big number. The part that deserves skepticism is extrapolation: jumping from "70+ platforms, 280 million installs" to "Agent Skills is already the de facto standard" skips over the third-party statistics, retention rates, and depth of actual usage that nobody has published.
Timeline: From a Claude Code Feature to a Neutral Foundation
Agent Skills has traveled a more complete path than most open standards, and it is worth replaying in full.
October 16, 2025: Agent Skills launches alongside Claude, available the same day in Claude apps, Claude Code, and the API (with the API beta header skills-2025-10-02). Anthropic later revealed that Skills were conceived during the development of Claude Code — solve your own agent's knowledge-reuse problem first, then think about standardization. Within weeks of launch, developers had published thousands of skills on GitHub and had written loaders for Codex, Cursor, and OpenCode unprompted. At this stage it was still a Claude feature, but the ecosystem had already slipped past the fence.
December 18, 2025: Agent Skills is released as an open standard, with an independent GitHub organization and the agentskills.io domain going live, along with organization-level skill management and a partner skills directory announced the same day. Later that month, Anthropic donates MCP to the Linux Foundation and the AAIF is established. October 22, 2026: the Agent Skills donation lands in full. Read in sequence, the path is clear: ship a product feature, extract it into an open standard, then hand over governance of the standard itself — name, logo, GitHub organization, domain — to a foundation.
For readers new to Agent Skills, some background. A skill is a folder whose core is a SKILL.md file: YAML frontmatter declaring a name and description, followed by Markdown task instructions, with optional scripts/, references/, and assets/ subdirectories for code and supporting material. Loading follows a three-tier progressive disclosure design: at startup only the name and description metadata are preloaded; the full SKILL.md is read when relevant; additional files are read only when needed — all to control context consumption. Its relationship to MCP is complementary, not competitive: MCP connects agents to external tools and data, while Skills package the procedural knowledge of how to do things.
What Handing Over Governance Actually Means
The most direct change is ownership of the standard. The official wording of the donation scope is unambiguous: the Agent Skills name and logo, the GitHub organization, and the agentskills.io domain. Neutrality now has an institutional guarantee — no single company, including Anthropic itself, can unilaterally steer the standard; changes must pass through the foundation's neutral governance process. The maintainer list confirms the spirit: "Engineers from Anthropic, Google, OpenAI, and other companies will maintain the standard as individuals," with David Soria Parra and Mahesh Murag as Lead Maintainers. One detail deserves care here. In Anthropic's announcement, David Soria Parra is identified as Agent Skills' lead maintainer; in the Linux Foundation press release quote, his title reads "Co-Creator and Lead Maintainer of MCP, Anthropic." Same person, two official phrasings — distinguish the context when quoting. His line, "Today, we're donating Agent Skills to the AAIF to ensure that anyone building with skills can shape what's next," appears in both documents.
For developers, the expectation is explicit: you can continue to use the standard or contribute to it without joining the AAIF. A roadmap has already been proposed — versioned releases, a formal proposal process, and security guidance for people who write and install skills. That last item may be the most consequential: when nearly 280 million installs mean skills get copied and executed at scale, security guidance is the most urgent piece of infrastructure this ecosystem lacks.
So what is this, really? Restrained read: ecosystem defense, not charity. For Anthropic, handing over the standard buys trust that accelerates adoption — third-party platforms face far less friction integrating a foundation-hosted neutral standard than integrating a competitor's product specification. The MCP donation already tested the logic: neutralization sped up diffusion rather than slowing it. Agent Skills is the same playbook executed a second time. But resist the urge to over-praise. What was donated is the standard and its governance, not the business: Anthropic still holds the platform-side skill ecosystem inside Claude, and standard neutrality does not mean competitive neutrality.
Open Questions and Cold Second Thoughts
Several uncertainties deserve honest disclosure. First, the list of AAIF's three launch projects — MCP, AGENTS.md, and goose — rests on consistent secondary reporting; the official press release confirms only the count, and this article has flagged it accordingly. Second, circulating details about SKILL.md character limits (a name capped at 64 characters, a description at 1024) come from secondary technical blogs; we have not verified them against the official specification on agentskills.io and do not endorse them here. Third, some outlets report that VS Code and ChatGPT integrated Skills within roughly 48 hours of December 18, 2025; this does not conflict with the official statement that 70-plus platforms integrated "since December 2025," but the precise dates lack official confirmation and are likewise not endorsed. Fourth, none of the three headline numbers has independent verification; no cross-checking source exists yet.
A colder thought: growing from 3 projects to 10 in a single year is a double-edged speed. Fast growth shows vendors genuinely want neutral stewardship, but whether governance bandwidth can keep pace with project count — and whether neutral process survives white-hot commercial competition — only time will tell. Announcing four simultaneous donations on one stage makes for good theater, but a foundation's real exam is not on launch day. It comes the first time the stakeholders disagree about where a standard should go.
One step further: what does any of this change for the average developer? In the short term, almost nothing. The standard is unchanged, SKILL.md is still SKILL.md, existing skills keep working, and the agentskills.io domain sits where it always did. What changed is beneath the waterline — the authority to modify the specification has moved from one company's repository into a foundation process, and future version evolution will be decided by many parties together. For anyone deciding whether to build a business on this standard, foundation stewardship answers the scariest question: will the standard be held hostage or abandoned by a single vendor? That is a heavyweight vote in any adoption decision. Whether the skill ecosystem flourishes as a result is a question for the coming months — watch how fast third-party platforms ship integrations, and whether a cross-platform author community emerges in the skills market. Those data points do not exist yet, and it would be premature to call the verdict.
What This Article Covers — and What It Doesn't
A final note on scope. This article is about governance: who donated what, to which foundation, and how the stewardship landscape changed. It is not a tutorial. If you want to learn how skills are actually built and used, the site has earlier coverage: for structured skills, read the book-to-skill piece on turning technical books into skills an agent can call on demand (linked below); for combining knowledge bases with skills, see the site's guide to AnythingLLM, a local knowledge base app. Published in the same batch are three infrastructure pieces — an open-source overview of AWS Infinity, a comparison of concurrent agent runtimes, and an SOP for the Claude Code sandbox runtime — which correspond to the execution-environment topics touched on above; one sentence here is all they get.
Sources
- Anthropic donation announcement (Product, Oct 22, 2026): https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation
- Linux Foundation press release (SAN JOSE dateline, Oct 22, 2026): https://www.linuxfoundation.org/press/agentic-ai-foundation-announces-four-new-projects-agent-skills-open-responses-openenv-and-opensandbox
- Agent Skills launch announcement (Oct 16, 2025): https://www.anthropic.com/news/skills
- Open standard release and organization-level skills (Dec 18, 2025): https://claude.com/blog/organization-skills-and-directory
- Anthropic engineering blog (SKILL.md structure and progressive disclosure): https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
- MCP donation and AAIF founding announcement (Dec 2025): https://anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
Related reading