Anthropic hands Agent Skills to the Linux Foundation: the skills standard moves under neutral governance, just as MCP did before it

Anthropic hands Agent Skills to the Linux Foundation: the skills standard moves under neutral governance, just like MCP before it

🔎 The format that won changes hands

On October 22, 2026, on stage at AGNTCon and MCPCon, Anthropic made official the donation of Agent Skills to the Agentic AI Foundation (AAIF), hosted by the Linux Foundation. Name, logo, GitHub organization, agentskills.io domain: everything moves under neutral governance. One year after MCP, Anthropic is replaying exactly the same playbook — and this time, nobody can claim it was an isolated PR stunt.

Why now? Because Agent Skills no longer belongs to Anthropic in practice. The format is integrated by more than 70 platforms, from ChatGPT to Cursor, including GitHub Copilot and VS Code. One million published skills, roughly 280 million installations. When your format becomes an industry's lingua franca, keeping it under exclusive control becomes a strategic liability, not an asset.

My take: this is the second act of a deliberate strategy — win the standard, then neutralize it to maximize its adoption. What remains is to dissect what this neutrality concretely changes for skills developers, what it locks in, and what it doesn't solve. That is the subject of this article.

The Essentials

• The donation: Anthropic transfers Agent Skills — name, logo, GitHub organization, and the agentskills.io domain — to the Agentic AI Foundation, announced on October 22, 2026, at AGNTCon + MCPCon.
• The precedent: this is Anthropic's second major donation after MCP, which was ceded to the same foundation upon its creation in late 2025. The same sequence: massive adoption first, neutralization of intellectual property afterward.
• The adoption: more than 70 integrating platforms (ChatGPT, Cursor, GitHub Copilot, VS Code), 1 million skills published, roughly 280 million installations (Anthropic figures, October 2026).
• The governance: individual maintainers from Anthropic, Google, and OpenAI, led by David Soria Parra (Anthropic) and Mahesh Murag (OpenAI). Versioned roadmap, public proposal process, security guidance integrated into the standard.
• The context: the AAIF has grown from 3 projects at its launch in late 2025 to 10 projects a year later, with the arrival of Open Responses (OpenAI), OpenEnv (Hugging Face), and OpenSandbox (Alibaba).

Recommended Tools

Note: Open Responses, OpenEnv, and OpenSandbox were just announced on October 22; the Linux Foundation press release remains the most reliable reference until their dedicated sites are up.

What Anthropic is actually giving away — and why it's not symbolic

Anthropic isn't "sharing" Agent Skills: it is transferring ownership of it. The name, the logo, the GitHub organization, and the agentskills.io domain all pass entirely under the control of the AAIF, as detailed in Anthropic's official announcement. Concretely, the issues, pull requests, and releases of the standard are now hosted in an organization that neither Anthropic nor any other vendor controls.

That's the difference between an open standard on the honor system and an open standard on paper. As long as the "Agent Skills" brand remained Anthropic's property, every integrating platform was legally dependent on a direct competitor. Today, the brand is held by a neutral foundation — the model that has proven itself with Kubernetes at the CNCF. No one can pull the legal rug out from under adopters.

Simon Willison immediately highlighted the symmetry with MCP in his October 22 post: same sequence, same destination, same logic. The only difference is that Agent Skills is younger — one year old at the time of the donation, versus a year and a half for MCP.

The MCP playbook, replayed exactly

Let's rewind. November 2024: Anthropic launches MCP, a protocol for connecting agents to tools and data. Adoption explodes, OpenAI and Google get on board. Late 2025: Anthropic hands MCP over to the Linux Foundation, which creates the Agentic AI Foundation for the occasion. October 2026: Agent Skills, same trajectory, same destination.

The pattern is legible. Anthropic opens a standard once it dominates, and converts a competitive advantage into industry infrastructure. Cynics will see it as PR. I see the most effective standardization strategy in the AI market — provided the governance keeps its promises. That's precisely what needs to be verified, and we're getting to it.

70+ platforms, 1 million skills: the adoption that made the donation inevitable

This donation is not an ideological choice, it's an arithmetic consequence. Agent Skills, launched by Anthropic in October 2025 in Claude (Claude Code, application, API) before being opened up as a standard, became within twelve months the de facto exchange format for agent skills: more than 70 platforms implement it, with 1 million skills published and roughly 280 million installations (figures announced by Anthropic, October 2026).

The strength of the format lies in its simplicity. A folder, a SKILL.md with metadata, optional scripts and references, and a progressive disclosure mechanism that only loads into context what the agent needs at the moment it needs it. Nothing in there is tied to a model. The same skill works with Claude Opus 4.7, GPT-5.5 or Gemini 3 Pro Deep Think — that's what made the format unavoidable, and our comparison of the best LLMs for AI agents confirms it.

The community ecosystem did the rest. The skill repositories flooding GitHub trending — we analyzed them in our article on the "skills" invading GitHub trending — proved that a market for skills existed beyond the publishers' catalogs.

And that's where ownership becomes a problem. When your competitors — OpenAI foremost — build features on your format, keeping the name and the domain under your exclusive control sends a signal of dependency that slows adoption. The donation removes that friction in one stroke. If you're looking for agents that already leverage this format on a daily basis, our selection of the best autonomous AI agents does the sorting.

AAIF: from 3 to 10 projects in one year, the control tower for agent standards

With this announcement, the Agentic AI Foundation consolidates every layer of the agentic stack under one neutral roof. Created in late 2025 with three founding projects — including MCP — the foundation is announcing ten projects one year later, according to the Linux Foundation press release.

The four newcomers sketch out a fairly complete map of the stack:

• Agent Skills (Anthropic): procedural know-how, what the agent knows how to do.
• Open Responses (OpenAI): the API layer, how you dialogue with agents.
• OpenEnv (Hugging Face): standardized environments for training and evaluating agents.
• OpenSandbox (Alibaba): sandboxed execution of agent-generated code.

Add MCP for tool connectivity, and you get a complete agent stack — skills, API, environments, execution, connections — entirely under neutral governance. Every brick has its standard, every standard has its foundation.

Notable fact: head-on competitors — Anthropic, OpenAI, Google, Alibaba, Hugging Face — are donors or maintainers of the same projects. The industry has made its call: at the infrastructure layer, neutral standards beat lock-in. This is the CNCF moment for agentic AI. In three years, I bet we'll look at the AAIF the way we look today at the foundation behind Kubernetes: invisible when everything works, indispensable as soon as you scratch the surface.

"Vendor-neutral": what it concretely changes for skill developers

Three concrete changes, not promises: a versioned, public roadmap, an open proposal process, and security guidance maintained at the standard level. Add to that a far-from-negligible psychological shift: you can invest weeks into a skill library without fear that a vendor will pivot the format out from under your feet.

A versioned roadmap, no longer a product whim

Before, the spec's evolution followed Anthropic's product rhythm. Now, the roadmap is versioned and published: you can pin a version, anticipate deprecations, plan your migrations. For anyone maintaining skills in production, that's the difference between a standard and a feature.

An open proposal process — and multi-vendor maintainers

A governance detail often glossed over: the maintainers are individuals, not corporate seats. David Soria Parra (Anthropic) and Mahesh Murag (OpenAI) lead the project, with maintainers coming from Anthropic, Google, and OpenAI. The engineer who designed SKILL.md at Anthropic and the one who integrated it into ChatGPT at OpenAI now maintain the same spec. Anyone can propose an evolution via the public proposal process. That, concretely, is what "vendor-neutral" means.

Security treated as a standards issue

A SKILL.md is a set of instructions an LLM will follow — hence a prompt injection attack surface, and a risk of malicious skills. Until now, each platform handled this problem its own way. Now, the security guidance is part of the standard, maintained collectively: audit expectations for skill publishers, sandboxing best practices, vulnerability disclosure processes.

To gauge what this packaging makes possible, revisit our breakdown of the Hermes agent's skill system: an agent that acquires capabilities by stacking skills — exactly the scenario that the standard's vendor-neutrality now makes durable.

What Neutrality Doesn't Solve

The governance handover doesn't level the ecosystem. Anthropic keeps most of the commercial value: the Claude models, Claude Code, the API. The spec is neutral; the best reference implementation is not. It's the classic standard-setter gap — Google with Chrome, Anthropic with Claude Code — and it's a perfectly legal moat.

Second limitation: "Agent Skills compatible" covers a very wide spectrum of implementations across the 70+ platforms. Some handle scripts, others don't; some respect progressive disclosure, others load everything. Portability is progressing, but it isn't magic — test before you promise.

Third limitation, less flattering: governance by committee is slow. A versioned roadmap and a proposal process mean predictability, but they also mean friction. Anthropic iterated faster on its own. That's the price of neutrality, and it's better to own it than to deny it.

Finally, a million published skills above all means an ocean of noise. The governance of the standard solves neither quality, nor discovery, nor trust in third-party skills. Those problems remain fully intact — and they may matter more than the ownership question.

Where to start if you're publishing skills today

Nothing changes in the format on October 22 — and that's exactly the good news. The spec has been stable for a year, the transfer breaks nothing, and the versioned roadmap guarantees it won't break anything without warning. The best time to publish your skills is now.

The basics that matter:

1. Read the spec on agentskills.io — it's now the neutral reference, no longer a vendor's blog.
2. Polish the description field: it's what the model reads first to decide whether to load your skill. Vague = invisible.
3. Test on at least two platforms (for example Claude Code and Cursor) before declaring your skill portable.
4. Treat security as a feature: document side effects, never require blind execution, follow the standard's security guidance.

On the model side, today's best agents — powered by Claude Opus 4.7, GPT-5.5, or Gemini 3 Pro Deep Think — follow well-written skills remarkably well. Quality comes from the precision of your instructions, not the price of the model. And for inspiration from mature use cases, look at what a personal-memory agent like OpenHuman makes possible: personal context, too, is becoming a packageable asset.

Last point, infrastructure: if your agents need to consume your skills continuously, a VPS like Hostinger is more than enough to get started (pricing to be checked on their site, October 2026). To run 100% locally with open source models, follow our guide on open source AI agents with Ollama.

❌ Common Mistakes

Mistake 1: confusing "neutral governance" with "community governance"

Neutrality means no vendor holds the intellectual property — not that everyone votes on everything. Maintainers remain gatekeepers with real power over merges. If you want to influence the roadmap, the path goes through the proposal process, not through social media.

Mistake 2: believing that all skills become portable by magic

The standard is neutral, the implementations are not. Some platforms implement subsets of the spec. The solution: identify your target platforms, test your skills on them, and avoid vendor-specific extensions in your critical skills.

Mistake 3: underestimating the attack surface of a SKILL.md

A skill is a set of instructions an agent will follow — a malicious skill is a hijacked agent. Never install a third-party skill without reading it, check its scripts and dependencies, and apply the security guidance now built into the standard. The neutralization of the format mechanically increases the number of skills in circulation — and therefore the number of bad ones.

Mistake 4: waiting for "the official version 1.0" to get started

The format has been in production for a year across dozens of platforms. The governance is changing, not the spec. Waiting means giving your skills a one-year delay — and your competitors a one-year head start.

❓ Frequently Asked Questions

Does Agent Skills still belong to Anthropic?

No. Since October 22, 2026, the name, the logo, the GitHub organization, and the agentskills.io domain belong to the Agentic AI Foundation, hosted by the Linux Foundation. Anthropic remains involved through its individual maintainers and its reference implementation, but no longer controls the standard's intellectual property.

What's the difference between Agent Skills and MCP?

They're complementary, not competing. MCP connects agents to external tools and data: APIs, databases, files. Agent Skills packages procedural know-how: instructions, scripts, and references that an agent loads when the situation calls for it. A well-designed agent uses both — MCP to connect, skills to know what to do.

How much does Agent Skills cost?

Nothing. The standard is open source and free (as of October 2026): the specification, the GitHub repository, and the documentation are freely available, including for commercial use. You only pay for the tokens of the models running your agents and, where applicable, your hosting. No license, no royalty.

Will my skills work in ChatGPT or Cursor?

That's the whole point of the standard: yes, in broad strokes. More than 70 platforms implement it, including ChatGPT, Cursor, GitHub Copilot, and VS Code. But implementations vary — some ignore scripts or progressive disclosure. Test your critical skills on your target platforms before any deployment.

Who decides on the Agent Skills roadmap now?

The project's individual maintainers within the AAIF, led by David Soria Parra (Anthropic) and Mahesh Murag (OpenAI), with contributors from Anthropic, Google, and OpenAI. Changes go through a public proposal process and a versioned roadmap. No more unilateral direction changes dictated by a product schedule.

Why the Linux Foundation rather than another organization?

Because it has already been home to MCP since the AAIF was created in late 2025, and the foundation has proven this model with Kubernetes and the CNCF. It holds the trademarks, provides the neutral legal framework, and brings governance experience. For Anthropic, it's the most credible shortcut to verifiable neutrality, not merely declared neutrality.

✅ Conclusion

By donating Agent Skills to the Linux Foundation one year after MCP, Anthropic loses nothing: it locks in its role as architect of the agentic standardization layer, while the foundation neutralizes what needs to be neutralized. For developers, the message is simple — the format is stable, neutral, and everywhere, so the best time to publish your skills is now. To take action, follow our guide to build an AI agent that leverages them end to end.