Cross-Tool Plugin Manifest Comparison

Addendum to Codex Manifest Shape Report | June 26, 2026
Sources: OpenAI Codex docs, Anthropic Claude Code docs, Cursor official plugins repo, OpenCode docs, HuggingFace Context Course
Platforms compared: Codex, Claude Code, Cursor, OpenCode, Pi

Platform Architecture Models

The AI coding tool ecosystem splits into two fundamental approaches:

Side-by-Side Manifest Comparison

Codex .codex-plugin/plugin.json

{
  "name": "text-processor-plugin",
  "version": "1.0.0",
  "description": "Text analysis skills",
  "skills": "./skills/",
  "mcpServers": "./.mcp.json",
  "apps": "./.app.json",
  "hooks": "./hooks/hooks.json",
  "interface": {
    "displayName": "Text Processor",
    "shortDescription": "Analyze text with AI",
    "category": "Productivity",
    "capabilities": ["Read", "Write"],
    "defaultPrompt": ["Analyze this text for readability"],
    "brandColor": "#10A37F",
    "logo": "./assets/logo.png"
  }
}

Claude Code .claude-plugin/plugin.json

{
  "name": "text-processor-plugin",
  "version": "1.0.0",
  "description": "Text analysis skills",
  "author": { "name": "Your Name" }
}

Notable: Claude Code's manifest is much leaner — components are discovered by convention (directory presence), not pointed to by explicit paths. No interface object.

OpenCode (No Manifest — Code is the plugin)

// .opencode/plugins/text-processor-plugin.ts
import type { Plugin } from "@opencode-ai/plugin"

export const TextProcessorPlugin: Plugin = async ({ project, client, $, directory }) => {
  return {
    "tool.execute.before": async (input) => {
      if (input.tool === "read") {
        await client.app.log({ body: { message: "Consider text analysis" } })
      }
    }
  }
}

Pi package.json

{
  "name": "text-processor-plugin",
  "version": "1.0.0",
  "keywords": ["pi-package"],
  "pi": {
    "skills": ["./skills"],
    "extensions": ["./extensions"],
    "prompts": ["./prompts"],
    "themes": ["./themes"]
  }
}

Directory Structure Comparison

Codex

my-plugin/
├── .codex-plugin/
│   └── plugin.json          ← Manifest (ONLY file here)
├── skills/
│   └── analyze-text/
│       └── SKILL.md
├── .mcp.json                ← MCP server config
├── .app.json                ← App/connector config  
├── hooks/
│   └── hooks.json
└── assets/
    ├── logo.png
    └── screenshot-1.png

Claude Code

my-plugin/
├── .claude-plugin/
│   └── plugin.json          ← Manifest (ONLY file here)
├── skills/
│   └── analyze-text/
│       └── SKILL.md
├── agents/                  ← UNIQUE: agent definitions
│   └── reviewer.md
├── .mcp.json                ← MCP server config
├── .lsp.json                ← UNIQUE: LSP server config
├── hooks/
│   └── hooks.json
├── monitors/                ← UNIQUE: background monitors
│   └── monitors.json
├── themes/                  ← UNIQUE: color themes
│   └── dracula.json
└── README.md

OpenCode

my-project/
├── .opencode/
│   ├── plugins/
│   │   └── text-processor.ts   ← Plugin IS the code
│   └── package.json
├── opencode.json               ← npm plugin references
└── README.md

Pi

my-package/
├── package.json                ← Manifest (no hidden dir)
├── skills/
│   └── analyze-text/
│       └── SKILL.md
├── extensions/                 ← Runtime TS/JS extensions
│   └── text-processor.ts
├── prompts/
└── themes/

Manifest Field Comparison

Component Types Supported

Distribution Models

Hook/Event Systems

Key Architectural Differences

1. Explicit Pointers vs Convention Discovery

Codex requires explicit manifest pointers:

{ "skills": "./skills/", "mcpServers": "./.mcp.json" }

Claude Code auto-discovers by directory presence — if skills/ exists at plugin root, skills are loaded. The manifest only needs name. This makes Claude Code plugins simpler to create but harder to reason about (implicit behavior).

2. Interface Metadata: Codex-Only

Codex has a rich interface object for marketplace presentation (brand colors, logos, screenshots, starter prompts). Claude Code has no equivalent in the manifest — all presentation metadata lives in the marketplace entry, not the plugin itself.

Why it matters: Codex plugins are self-describing for marketplace purposes. Claude Code plugins depend on their marketplace entry for discoverability.

3. App Connectors: Codex-Exclusive

Only Codex has .app.json for third-party service authentication (OAuth, API keys). Claude Code, OpenCode, and Cursor don't bundle auth connectors in their plugin systems — they expect MCP servers to handle external service integration.

Why it matters: Codex treats "connecting to Slack/GitHub/Figma" as a first-class plugin concern. Others offload this entirely to MCP.

4. Subagents: Claude Code-Exclusive

Only Claude Code supports agents/ directories with agent definitions (model selection, tool restrictions, isolation). Codex has no agent bundling — its "subagent" concept operates outside the plugin system.

5. LSP + Monitors + Themes: Claude Code's Breadth

Claude Code has the richest component surface: LSP servers (real-time code intelligence), monitors (background stdout watchers), and themes. These make Claude Code plugins closer to full IDE extensions. Codex is narrower but cleaner.

6. Validation Philosophy

This is a significant divergence: Codex plugins can carry forward-compatible experimental fields. Claude Code's strict validation means older tools break on newer manifest features.

7. Cross-Tool Portability

MCP is the universal interop layer. A well-built MCP server works across all tools. Everything above MCP (skills, hooks, marketplace metadata) is tool-specific — but the functional logic (the MCP server code) is write-once, run-anywhere.

Summary: When to Use Which Model

The Convergence Thesis

Despite surface differences, the ecosystem is converging:

1. MCP as universal glue — All tools consume MCP servers. This is the write-once layer.
2. Skills as a shared concept — SKILL.md with YAML frontmatter appears in Codex, Claude Code, and Pi. The format is nearly identical.
3. Git-native distribution — Both Codex and Claude Code use marketplace.json catalogs pointing at Git repos. No compilation step, no build system.
4. Plugin = directory — Both manifest-first platforms share the same physical pattern: a hidden config dir (.codex-plugin/ or .claude-plugin/) with plugin.json, and all components at the plugin root.

The key difference is scope of ambition: Codex focuses on Skills + MCP + Apps (clean, portable). Claude Code maximizes component variety (agents, LSP, monitors, themes). OpenCode goes programmatic (full JS/TS control). But MCP underneath means the functional tools are portable regardless of which wrapper you choose.

Sources: OpenAI Codex Plugins, Claude Code Plugins Reference, Cursor Plugin Ecosystem, OpenCode Plugins, HuggingFace Context Course Unit 3