Every skill on shareskills is now an Agent Plugins 1.0 package

On August 6, a coalition including Amazon, Microsoft, OpenAI, Vercel, Cursor and Google published Agent Plugins 1.0 — a vendor-neutral package format for the components that extend AI agents, launching with support in VS Code, GitHub Copilot, Cursor, ChatGPT and Kiro. The full specification is on GitHub, and it is a genuinely tidy piece of work: a closed `plugin.json` manifest, a `skills/` directory, an optional `mcp.json`, and very little else.

As of today, every skill on shareskills downloads as a conformant Agent Plugins 1.0 package — a "Download as plugin" button on each directory page, or one predictable URL. We validate what we emit against the coalition’s own published JSON Schema, not our reading of it.

What you get

The download is a zip that unpacks to exactly the layout the spec prescribes — the manifest at the root, the skill in `skills/`:

handy-skill/
├── plugin.json
└── skills/handy-skill/
    ├── SKILL.md
    └── references/…

The manifest carries the skill’s name, description, version and publisher — plus a provenance block under our own extension namespace, which conformant clients ignore and humans can read:

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "handy-skill",
  "version": "4.0.0",
  "description": "One-line description of what triggers this skill.",
  "author": { "name": "Example Publishing Co" },
  "homepage": "https://shareskills.ai/directory/example-co/handy-skill",
  "extensions": {
    "ai.shareskills": {
      "source": "https://shareskills.ai/directory/example-co/handy-skill",
      "team": "example-co",
      "native_version": 4,
      "exported_at": "2026-08-10T08:51:30Z"
    }
  }
}

Why there was nothing to convert

The spec’s `skills/` directory holds Agent Skills — a `SKILL.md` with supporting folders. That has been shareskills’ storage format since day one, byte for byte. Your library was already the payload; the standard just gave it a portable envelope. Which means this works retroactively: every skill anyone has ever published here exports today, with no action from its author.

The hole in the spec is the point

Read the specification and notice what it declines to define: installation, distribution, provenance verification, permission models, and marketplaces are all explicitly out of scope. That is not an oversight — it is disciplined scoping, and it is also an honest admission that the hard part of sharing agent capabilities was never the folder layout.

We wrote a whole post about that hard part the week before this standard landed: distribution is an attack surface, updates are how trust gets exploited, and somebody has to be accountable for the layer between “a package exists” and “I should run this.” The standard defines the package. The store — identity, permissions, updates you can trust — is the part we build.

One honest caveat: exports are snapshots

A downloaded plugin leaves the shareskills update channel. It will not receive new versions, and the update gate that reviews risky changes does not travel with it — it is frozen bread, exactly as fresh as the day you exported it. The `ai.shareskills.source` URL in the manifest is the way home: check it when you want the current version. Agents connected over MCP should keep using the live channel, which stays gated and updateable.

Try it

• Open any skill in the public directory and hit Download as plugin.
• Or fetch it directly: `/api/plugin-export/<team>/<skill>` returns the zip.
• Unpack it into wherever your client keeps plugins — VS Code, Copilot, Cursor, ChatGPT and Kiro all read the format.

Interoperability is good for everyone shipping skills, and we would rather meet the ecosystem where it is than wall a garden. The package is standard now. Make the store earn you.