A plugin is a package of tools and skills that installs into any harness.

What it is

A plugin is one named, versioned package of MCP servers and skills. Install it into a harness and every task of that harness gets its tools and skills beside the harness's own, without copying them into each configuration.

Plugins use the open Agent Plugins format, so the same package installs into any harness that takes plugins, and a harness you have already configured can be exported as one.

What a package holds

plugin.json

The manifest, at the root, and required. It names the Agent Plugins schema it follows, and its name names the plugin: 1 to 64 lowercase letters, digits, hyphens and dots, starting and ending with a letter or digit.

mcp.json

MCP servers over stdio, streamable HTTP or SSE. A remote server is reached over https.

skills/<name>/SKILL.md

One folder per skill. The skill's name in its frontmatter matches the folder name, and its description is at most 1,024 characters.

Anything else

Scripts, templates and data the skills and servers use. Every file directly in bin/, and each .sh file directly in hooks/, is made executable.

Installing one

• In the console, open a harness's settings and choose Install from folder under Plugins. The package installs when you save, and each installed plugin can be downloaded as a zip or removed.
• From the API, send the package's files in the harness's plugins list when you create or update it. A text file goes in content and a binary one in content_b64.
• To keep an installed package as it is, send it back as {"name", "blob"} the way the harness returned it. Sending files replaces it, and leaving it out of the list uninstalls it.

POST /v1/harnesses
{
  "name": "Contract Review Agent",
  "base": "claude-code",
  "plugins": [{
    "enabled": true,
    "files": [
      {"path": "plugin.json", "content": "{\"$schema\": \"https://agent-plugins.org/schemas/1.0.0/plugin.schema.json\", \"name\": \"contract-review\", \"version\": \"1.0.0\"}"},
      {"path": "skills/risk-checklist/SKILL.md", "content": "---\nname: risk-checklist\ndescription: ...\n---\n..."}
    ]
  }]
}

What a task gets

Each enabled plugin adds its skills and MCP servers to the harness's own. The package is unpacked into the task under .harness/plugins/, with a writable folder for it under .harness/plugin-data/, each named after the plugin with hyphens and dots written as underscores, and stdio servers find the two through PLUGIN_ROOT and PLUGIN_DATA.

A server or skill that cannot load does not fail the harness. The harness record lists it, with the reason, under skipped or notLoaded, so you can see what a task will and will not have.

Exporting and reading back

GET /v1/harnesses/{id}/plugin                  the harness's own tools and skills as a package, credentials left out
GET /v1/harnesses/{id}/plugins/{name}/files    the files of one installed package

Errors

422 plugin_invalid

The package or its manifest is not valid, for example a missing plugin.json or a path outside the package.

422 unsupported_plugin_schema

The manifest names a schema this server does not read. The error lists the supported ones.

413 file_too_large

The package is over the size limit.

409 plugin_conflict

An MCP server or skill name clashes with the harness or another plugin, or two plugins share a name.

404 plugin_not_found

No installed plugin has that name.

Limits

One package

Up to 16 MiB.

Manifest

Up to 24,000 bytes, with what is derived from it.

System One

Takes no skills, so a plugin's skills are left out on that base.