Universal AI Plugin Marketplace Template

Author AI assistant plugins once and distribute them to all major platforms at the
highest fidelity each one accepts.

Supported Platforms

Quick Start

Create a new plugin

pnpm run scaffold my-plugin

Validate all plugins

pnpm run validate

Build standalone exports

pnpm run build:standalone

This generates standalone directories in dist/ for platforms that require
repo-root manifests (Gemini CLI and Kiro). Claude Code, Cursor, and Codex read
plugins directly from plugins/<name>/ via their marketplace registries.

Repository Structure

ai-plugin-marketplace-template/
├── .claude-plugin/
│   └── marketplace.json              # Claude Code marketplace registry
├── .cursor-plugin/
│   └── marketplace.json              # Cursor marketplace registry
├── .agents/
│   └── plugins/
│       └── marketplace.json          # OpenAI Codex marketplace registry
├── plugins/
│   └── <plugin-name>/                # One directory per plugin
│       ├── .claude-plugin/
│       │   └── plugin.json           # Claude Code manifest
│       ├── .cursor-plugin/
│       │   └── plugin.json           # Cursor manifest
│       ├── .codex-plugin/
│       │   └── plugin.json           # OpenAI Codex manifest
│       ├── gemini-extension.json     # Gemini CLI manifest
│       ├── POWER.md                  # Kiro power entry point
│       ├── GEMINI.md                 # Gemini CLI context file
│       ├── .mcp.json                 # MCP config (Claude Code / Cursor / Codex)
│       ├── mcp.json                  # MCP config (Kiro)
│       ├── skills/                   # SKILL.md files (universal)
│       ├── agents/                   # Agent definitions (.md)
│       ├── rules/                    # Rules (.md for Claude, .mdc for Cursor)
│       ├── steering/                 # Kiro steering files
│       ├── commands/                 # Commands (.md for Claude/Cursor, .toml for Gemini)
│       ├── hooks/                    # Hook definitions (claude.yaml source)
│       ├── README.md
│       └── LICENSE
├── src/
│   ├── validate.ts                   # Validate plugins and manifests
│   ├── scaffold.ts                   # Create new plugin from template
│   ├── build-hooks.ts                # Convert hooks YAML → per-target JSON
│   └── build-standalone.ts           # Generate standalone exports
├── schemas/                          # JSON Schemas for plugin and marketplace manifests
├── templates/                        # Templates for scaffolding
├── dist/                             # Generated standalone repos
│   ├── gemini/
│   └── kiro/
└── README.md

Platform Compatibility Matrix

How It Works

Platform terminology note: Claude Code, Cursor, and Codex call these
plugins; Kiro calls them powers; Gemini CLI calls them extensions;
Vercel calls them skills. This template uses "plugin" as the generic term
throughout.

Each plugin is authored once under plugins/<name>/ as a superset of every
component type — agents/, skills/, commands/, hooks/, rules/,
steering/, and MCP config. The build pipeline emits the richest native
representation each target tool accepts. Components a given target cannot
express natively are simply absent from its output — they are not flattened
into SKILL.md as a fallback.

Tier 1: Rich plugins (Claude Code, Cursor, Gemini CLI, OpenAI Codex)

These platforms accept full plugin packages bundling multiple component types.

• Claude Code and Cursor share the same internal layout — skills/,
  agents/, hooks/, commands/, .mcp.json are identical. Only the
  manifest wrapper differs (.claude-plugin/plugin.json vs
  .cursor-plugin/plugin.json). Plugins are referenced directly from the
  per-tool root marketplace.json.
• Gemini CLI ships as an extension with a richer
  gemini-extension.json (MCP servers, contextFileName, excludeTools,
  settings) and native auto-discovered directories for skills/,
  agents/*.md, commands/*.toml, hooks/hooks.json, and policies/*.toml.
  pnpm run build:standalone writes the self-contained extension to
  dist/gemini/<name>/, with Claude-flavored agent tool names rewritten to
  Gemini equivalents (Read → read_file, Write → write_file, etc.) and
  the hook YAML converted to Gemini's hooks.json.
• OpenAI Codex reads plugins directly from this repo via
  ./.agents/plugins/marketplace.json, with each plugin's manifest at
  plugins/<name>/.codex-plugin/plugin.json. The Codex manifest carries an
  interface block (displayName, shortDescription, longDescription,
  category, capabilities, brandColor, etc.) used to render the plugin in
  the Codex UI. Codex natively supports skills, MCP, apps, and interface
  metadata — sub-agents, commands, and hooks are absent from Codex output
  (they remain available to the Claude/Cursor/Gemini targets).

Tier 2: Standalone export (Kiro)

Kiro expects its manifest at the repository root. pnpm run build:standalone
copies each plugin into dist/kiro/<name>/ with POWER.md, steering/,
skills/, and auto-converted .kiro/agents/*.json configs (generated from the
Claude agent .md files). Kiro reads skills via steering context files rather
than natively parsing SKILL.md frontmatter, and has no commands concept.

Tier 3: Lossy fallback (Vercel Skills CLI)

npx skills add <owner>/<repo> auto-discovers every SKILL.md in the
repository. Nothing else — sub-agents, commands, hooks, rules, and MCP configs
are not surfaced to Skills CLI consumers. Use this tier when a plugin's
value is carried entirely by its skills.

OpenAI Codex

Codex plugins live alongside the other platforms' manifests:

• Per-plugin manifest: plugins/<name>/.codex-plugin/plugin.json.
• Repo-root marketplace: .agents/plugins/marketplace.json with entries of the
  form

  {
    "name": "<plugin-name>",
    "source": { "source": "local", "path": "./plugins/<plugin-name>" },
    "policy": { "installation": "AVAILABLE", "authentication": "ON_INSTALL" }
  }
  

• The interface block on the plugin manifest drives Codex UI rendering
  (displayName, shortDescription, longDescription, category,
  capabilities, brandColor, composerIcon, logo, screenshots,
  defaultPrompt, websiteURL, privacyPolicyURL, termsOfServiceURL).

After installing a Codex user source pointing at this repo, codex /plugins
will list the plugin with its interface metadata.

Example Plugin: skill-evaluator

The included skill-evaluator plugin demonstrates the full multi-platform
pattern end-to-end. It evaluates AI skills across model tiers (opus → sonnet →
haiku) using blind sub-agent testing.

See plugins/skill-evaluator/README.md for
details.

Creating a Plugin Manually

If you prefer to create a plugin without the scaffold script:

1. Create a directory under plugins/.
2. Add platform manifests (see the skill-evaluator plugin for the Tier 1
   shape).
3. Add your skills, agents, rules, commands, and hooks.
4. Update all three root marketplace files:
   • .claude-plugin/marketplace.json
   • .cursor-plugin/marketplace.json
   • .agents/plugins/marketplace.json
5. Run pnpm run validate to verify everything is correct.
6. Run pnpm run build:standalone to generate Gemini and Kiro exports.

npm Scripts

License

MIT