Skip to content

Plugins for other agents

Last reviewed October 24, 2026

Every agent that loads plugins wants the same thing underneath: a manifest that names the plugin, and a skills/ folder beside it. They disagree about what the manifest is called and where it goes. Since a plugin can carry more than one manifest, one folder can serve most of them at once.

The Agent Plugins format

Agent Plugins 1.0 is an open format for a plugin that is not tied to one vendor (Agent Plugins specification). The manifest is a plugin.json at the plugin root with two required fields, a $schema and a lowercase name. Skills go in skills/ and MCP servers in mcp.json, with no leading dot. Settings for one agent go under extensions, keyed by a reverse-domain name such as com.openai, and agents skip the keys they do not know. The format covers the plugin only. It says nothing about marketplaces or installing.

plugin.json
{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
  "name": "release-notes",
  "version": "1.0.0",
  "description": "Drafts release notes from the git history since the last tag.",
  "author": { "name": "Your Name", "email": "you@example.com" },
  "license": "MIT",
  "repository": "https://github.com/your-name/release-tools"
}

This manifest passes the published schema at https://agent-plugins.org/schemas/1.0.0/plugin.schema.json. Codex, Copilot CLI and Cursor read it.

One folder for Claude Code, Codex, Copilot CLI, Cursor and Gemini CLI

Claude Code's docs do not say it reads a root plugin.json, so keep its own manifest as well. Add a Gemini CLI manifest and the same folder covers five agents:

layout
release-notes/
├── plugin.json                  Codex, Copilot CLI, Cursor
├── .claude-plugin/plugin.json   Claude Code
├── gemini-extension.json        Gemini CLI
└── skills/
    └── writing-release-notes/
        └── SKILL.md             all of them

The example from the build step, with these three manifests, passes claude plugin validate --strict, the Agent Plugins schema, gemini extensions validate and agy plugin validate. Copilot CLI, Codex, Gemini CLI and Antigravity each installed it and found its skill.

The shared folder stops at skills. Two parts need care:

  • MCP servers. Claude Code reads .mcp.json, the Agent Plugins agents read mcp.json, and Gemini CLI wants them in its manifest. No single file is documented to work in all of them.
  • Hooks. Each agent looks for them in its own place and runs them on its own events. Antigravity reads only hooks.json at the plugin root, so it skipped the example's hooks/hooks.json, and Copilot CLI reads hooks for an Agent Plugins plugin from com.github.copilot/. Test a hook in every agent that will load it, or leave hooks out of a plugin meant for all of them.

Codex

Codex reads the Agent Plugins plugin.json and keeps its own settings under extensions.com.openai. An older .codex-plugin/plugin.json still works as a fallback. Codex holds back a plugin's hooks until the user has reviewed and trusted them, and a plugin with hooks cannot go in OpenAI's public directory (Codex plugins).

People add a marketplace with codex plugin marketplace add and install with codex plugin add. Codex 0.164 also installed a plugin that had only Claude Code's manifest, took its version from it and found its skill.

Copilot CLI

Copilot CLI uses a root Agent Plugins manifest when there is one. Without it, it looks through older locations ending with .claude-plugin/plugin.json, so a Claude Code plugin loads as it is. In Agent Plugins mode it reads its own agents, commands, rules, hooks and language servers from a com.github.copilot/ folder (Copilot CLI plugin reference). People install with copilot plugin install, from a marketplace or a repository but not a local folder. Given a plugin with only Claude Code's manifest, Copilot CLI 1.0.103 installed it and its skill.

Cursor

Cursor takes the Agent Plugins format or its own .cursor-plugin/plugin.json, where only name is required. Beside it go skills/, rules/, agents/, commands/, hooks/hooks.json and mcp.json. A variables field declares settings the user supplies, such as an API token for a server (Cursor plugins).

The Cursor CLI loads a plugin folder for one session with agent --plugin-dir, and adds a marketplace from a git address with agent plugin marketplace add. Both need a signed-in Cursor account. Loaded that way, Cursor CLI 2026.10.22 found the example's skill, both from the shared folder and from a plugin with only Claude Code's manifest.

Gemini CLI

Gemini CLI calls a plugin an extension. Its manifest is gemini-extension.json at the very root of the repository, and the docs ask for a lowercase name that matches the folder. MCP servers go in the manifest's mcpServers, with paths starting ${extensionPath}. Folders: skills/, commands/ (in TOML), agents/, hooks/hooks.json and policies/. A GEMINI.md in the folder is loaded as context (extension reference).

gemini-extension.json
{
  "name": "release-notes",
  "version": "1.0.0",
  "description": "Drafts release notes from the git history since the last tag."
}

Check it, then link the folder so Gemini CLI loads your working copy while you change it:

shell
$ gemini extensions validate ./release-notes
Extension ./release-notes has been successfully validated.
$ gemini extensions link ./release-notes

Antigravity

An Antigravity plugin has a plugin.json with a name and an optional description, and its own optional $schema, https://antigravity.google/schemas/v1/plugin.json. Beside it go skills/, agents/, rules/, hooks.json and mcp_config.json (Antigravity plugins). People install from a GitHub address or a local folder with agy plugin install.

The CLI does more than the docs say. agy plugin validate checks a plugin folder, and on the example it reported that an Agent Plugins manifest loads as an Agent Plugins plugin while Antigravity's spec flag is on, and as its own kind of plugin otherwise. Given a Claude Code plugin, agy plugin install converts it, hook included.

Building something else? See the guides to a skill and an MCP server.