Plugin

A plugin is a Git repository capable of sharing at once a bundle of components.

Plugins are usually bundled within marketplaces to be installed by agent runtimes, but they can also be installed directly through the Agent Package Manager or some agent runtimes (like goose).

Structure

A plugin is commonly (may change depending on the agent runtime) identified by a plugin.json file. Its place and content (with small variations) depend on the target agent runtime(s).

Format: Agent Plugins
Read by: Copilot, Cursor, Devin, Hermes
Environment variables:
- PLUGIN_ROOT (plugin directory)
- PLUGIN_DATA (plugin data directory)

repository/
├── plugin.json
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── mcp.json
└── com.example.client/
    ├── agents/
    │   └── reviewer.agent.md
    └── hooks/
        └── hooks.json

The com.example.client directory is a client extension. It contains components in an agent-runtime-specific format (hooks, MCP servers, agents, etc.).

Copilot: com.github.copilot/

Format: Claude Code
Also read by: Codex (partial), Copilot (partial), Devin (partial)
Environment variables:
- CLAUDE_PLUGIN_ROOT (plugin directory)
- CLAUDE_PLUGIN_DATA (plugin data directory)
- CLAUDE_PROJECT_DIR (workspace directory)
- CLAUDE_PLUGIN_OPTION_<KEY> (userConfig values, compatible with Claude Code only)

repository/
├── .claude-plugin/
│   └── plugin.json
├── agents/
│   └── reviewer.md
├── commands/
│   └── special.md
├── hooks/
│   └── hooks.json
├── monitors/
│   └── monitors.json
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── .mcp.json
└── .lsp.json

Format: Codex
Environment variables:
- PLUGIN_ROOT (plugin root directory)
- PLUGIN_DATA (plugin data directory)
- CLAUDE_PLUGIN_ROOT / CLAUDE_PLUGIN_DATA (aliases kept for compatibility)

repository/
├── .codex-plugin/
│   └── plugin.json
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── hooks/
│   └── hooks.json
├── .mcp.json
└── .app.json

Format: Copilot
Environment variables:
- PLUGIN_ROOT (plugin directory)
- COPILOT_PLUGIN_DATA (plugin data directory, aliased as CLAUDE_PLUGIN_DATA)

repository/
├── plugin.json
├── agents/
│   └── reviewer.agent.md
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── hooks.json
└── .mcp.json

Format: Cursor

repository/
├── .cursor-plugin/
│   └── plugin.json
├── rules/
│   └── rule-name.mdc
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── hooks/
│   └── hooks.json
└── mcp.json

Format: Antigravity

repository/
├── plugin.json
├── mcp_config.json
├── hooks.json
├── skills/
│   └── skill-name/
│       └── SKILL.md
├── agents/
│   └── reviewer.md
└── rules/
    └── rule-name.md

Format: Devin
Environment variables:
- PLUGIN_ROOT (plugin directory)
- PLUGIN_DATA (plugin data directory)
- DEVIN_PROJECT_DIR (workspace directory)

repository/
├── .devin-plugin/
│   └── plugin.json
├── AGENTS.md
├── rules/
│   └── rule-name.md
├── agents/
│   └── reviewer.md
├── hooks.json
├── .mcp.json
└── skills/
    └── skill-name/
        └── SKILL.md

Format: Hermes Agent

repository/
├── plugin.yaml
├── __init__.py
└── skills/
    └── skill-name/
        └── SKILL.md

# __init__.py
def register(ctx):
    ctx.register_hook("pre_tool_call", my_pre_tool_call)
    ctx.register_tool("my_tool", my_tool_handler)
    ctx.register_skill("my-skill", "skills/my-skill/SKILL.md")
    ctx.register_command("my-command", my_command_handler)
    # `register_cli_command`, `register_platform`, `register_memory_provider`
    # `register_context_engine`, `register_image_gen_provider`
    # `register_slack_action_handler`, `register_middleware`

Limitations

LSP format

While LSPs can be defined in .lsp.json, a preferable way is to define them inline in plugin.json because its format varies between agent runtimes.