Anthropic published a standard in December 2025. It is a folder with one text file in it, called SKILL.md. Anthropic says more than 70 04 agent platforms have built support for it since then, including ChatGPT, Cursor, GitHub Copilot and VS Code.
On October 22, 2026 10, Anthropic gave it away. Anthropic announced on October 22, 2026 11 that it is donating the Agent Skills name, logo, GitHub organization and agentskills.io domain to the Agentic AI Foundation, a Linux Foundation project. Anthropic says anyone can keep using the standard without joining.
Seventy platforms in ten months is about 7 new integrations a month. Several of those platforms belong to companies that sell agents against Claude every day.
Anthropic also says 1 million 12 skills have been published and nearly 280 million 13 installed. Those are the company's own counts. Deyron Labs, which covered the announcement, notes they were not independently verified. Nobody outside Anthropic has defined what an "install" means.
The proof has limits, and the direction is still clear. The company that wrote the format chose to own less of it, and its rivals ship support anyway. For anyone building agents, that changes where a capability should live.
Anthropic Trades Ownership for Default Status
The Neutral Interface Trade has now run twice at one company. In December 2025, Anthropic donated the Model Context Protocol to the newly formed Agentic AI Foundation. The Linux Foundation announced that body on December 9, 2025 16. Anthropic and Block were founding contributors. So was OpenAI. MCP set how agents reach tools and data. Agent Skills sets how they learn a task.
Ten months of adoption in four figures.
The trade sorts a company's assets into two piles. The first pile is title, meaning the brand itself. It also covers the repository and the domain. Anthropic handed all of it to a neutral owner. The second pile is everything that runs on top of the format. Anthropic kept that.
What stayed with Anthropic is real. Skills were born inside Claude Code, so Claude carries the deepest native support. Anthropic says engineers from Anthropic and Google will maintain the standard as individuals. So will engineers from OpenAI and other companies. David Soria Parra and Mahesh Murag are the lead maintainers. The origin company keeps its people in the room, along with its knowledge of how the format behaves at scale.
Rivals gain permission. When one vendor owns a format, every adopter has to ask whether supporting it creates dependence on a competitor. A Linux Foundation home removes that question for every host. That includes Microsoft and Google.
The pattern is spreading past Anthropic. Unite.ai reports the foundation added four projects on October 22, 2026 07, including OpenAI's Open Responses and Alibaba's OpenSandbox. Each contributor handed an interface to a neutral home and kept the product that runs on it.
Where the Cash Actually Lands
Users came first. Anthropic says it built skills while working on Claude Code, which handled general knowledge work well but often lacked the procedural knowledge real business tasks need. Within weeks of the October 2025 launch, developers published thousands of skills on GitHub and wrote loaders for Codex and Cursor. They wrote one for OpenCode as well, according to Anthropic. A loader is a small program that reads a skill folder and hands its contents to an agent.
The order of events matters. The market picked the format before Anthropic formally opened it in December 2025, so the October donation ratified a fact already on the ground. Anthropic went from launch to neutral governance in about a year. That is fast for a company giving up a brand.
Anthropic published the format as an open standard, and the foundation now owns the name. Divide Anthropic's nearly 280 million 13 installs by its 1 million 12 published skills and you get about 280 27 installs per skill. The average hides what matters. Most installs likely sit on a few popular skills, and an install says nothing about a finished task or a paid model call.
The cash, if it comes, shows up where skills run. Every time a skill executes on Claude, Anthropic sells inference. The bet is that a bigger pool of portable skills sends more complex work to whichever model runs them best. It's the old move: open the interface, then monetize the stack around it.
Now the weak points, stated plainly. A shared folder format guarantees that a skill loads in many hosts. It does not guarantee the same results, because hosts vary in context windows and system prompts. They also differ in tool schemas and safety policies. It is also unclear whether "integrated" means full production support or a weekend loader. Anthropic has not published a definition or platform-by-platform usage.
Skills can carry scripts, too, which makes them code with a friendly name. TechJournal advises developers to inspect every downloaded skill before installing it, because bundled instructions and code can create security risks. A popular skill loaded across dozens of hosts is a large target.
The evidence supports one decision today. Package repeatable agent work in the Agent Skills format, because the biggest developer surfaces already load it and a folder costs almost nothing to write. Two events would change that call. A major host could ship a private extension that skills start depending on. Or a high-profile malicious skill could push enterprises to ban third-party skills.
We think the donation is a sound trade for Anthropic and a good deal for builders. Our caveat is that the format is thin by design. The moat moves to whatever executes it well, such as evaluation and permissions. Data and trust belong on that list too.
The format is thin and execution decides
Competitors ship support anyway.
ChatGPT, Cursor, GitHub Copilot and VS Code all load SKILL.md folders. A Linux Foundation home removes the worry that supporting the format creates dependence on Anthropic.
Installs remain undefined.
Anthropic's install count works out to about 280 27 per skill. Nobody has defined an install, and Anthropic has not published platform-by-platform usage or what "integrated" means.
Skills can carry scripts.
TechJournal advises developers to inspect every downloaded skill before installing it. A popular skill loaded across dozens of hosts is a large target, and one malicious skill could push enterprises to ban third-party skills.
Skills Turn Into Plumbing by 2028
The foundation's own numbers show the speed. "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10 14," Manik Surtani said, according to Unite.ai. Surtani is the foundation's CTO and co-founder. That works out to seven new projects in about ten months.
On Anthropic's count, Agent Skills grew at about 7 platforms a month. Assume the pace halves to 3.5 29 a month as the easy adopters run out. That adds 42 30 platforms a year and puts support near 112 platforms by October 2027. Few markets hold that many agent hosts worth counting, so the curve flattens soon after. Then the format stops being news.
The flywheel is simple. More published skills give each host a reason to add support. More hosts give each skill author a bigger audience. Each turn makes the next adopter's decision easier. That is how a convention compounds into infrastructure.
MCP walked the same road a year earlier. By October 2026, the foundation hosted MCP and Agent Skills under one roof. OpenAI's AGENTS.md lived there too. The agent stack is becoming modular, one donated interface at a time.
For a builder, the risk math is lopsided. Writing a capability as a portable folder costs an afternoon. Hard-coding it into one vendor's prompt format costs a rewrite for every new host you want, and Anthropic counts more than 70 04 of them.
Once the format is plumbing, money moves to the layers around it, like registries that rank skills and scanners that check them. Enterprise controls that gate them are another. The richest skills may encode hard operational knowledge, such as insurance claims or tax procedures. The data is mixed on who captures most of that value. A five-year forecast here is a sketch at best.
Port One Skill Across Two Agents
Step one: pick a task your team repeats every week and that new people get wrong at first. A monthly client report or a code review checklist works well. Write the steps down in plain sentences, the way you would brief a sharp temp on day one.
Step two: create a folder named for the task. Inside it, add a SKILL.md file with a name and a description at the top. Those two fields are required metadata, meaning the short label an agent reads to decide whether the skill applies. Put any scripts and reference files in the same folder.
Step three: strip out anything tied to one host. Remove tool names that only one product knows and file paths you assumed. State the inputs, the outputs, the required tools and every point where a human must approve.
Step four: load the same folder into two hosts, such as Claude Code and VS Code with GitHub Copilot. Run the same five test inputs in each one. Log each run's outcome and duration in a simple sheet.
Step five: expect the two hosts to disagree. The disagreement is the useful finding. Fix the skill wherever your instructions were vague, and write down any host requirement you cannot remove.
Step six: tag the folder v0.1.0 and keep it in version control. A version tag stamps each release so a host can pin an exact copy. Before you install anyone else's skill, read every script inside it.
Rerun the five tests each month and whenever a host ships an update. Track the share of runs that finish correctly. Your completion count is the one adoption number you can verify yourself.
Port one recurring task into a portable skill.
- Write the task as a skill folder. Pick a weekly task new hires get wrong at first, then create a folder with a SKILL.md file that starts with the required name and description fields.
- Strip host-specific details. Remove tool names and file paths that only one product knows. State the inputs, outputs, required tools and every human approval point.
- Test across two hosts. Load the folder into Claude Code and VS Code with GitHub Copilot, run five identical test inputs in each, log outcomes and duration, then tag the folder v0.1.0 in version control.
Write the skill once and let every host load it.
Anthropic kept the product layer and gave the interface to a neutral home, and rivals followed with support. The format guarantees that a skill loads. It does not guarantee identical results, so the value moves to evaluation, permissions, data and trust. We think builders should package repeatable work as portable skills now, test them across hosts every month and read every script before installing one.