ChatGPT / Codex Extension Concepts

Summary Table

Layered Mental Model

Ordered by learning sequence

In a single sentence each:

• Slash Commands are the shortcuts — quick triggers for frequent workflows.
• Function Calling / MCP is the nervous system — connecting the model to external tools and real-world data.
• Skills are the playbook — modular instructions the agent discovers and loads on demand.
• Custom GPTs are the product layer — full personas with custom instructions, tools, and a storefront presence.
• Plugins are the shipping container — bundling skills, MCP servers, and apps for easy distribution (added in v0.110.0).

Installation Methods

⚡ Slash Commands (Codex CLI)

# Built-in commands ship with Codex
/resume        # reopen saved conversations
/fork          # clone conversation into new thread
/compact       # summarize to conserve tokens
/clear         # start a fresh chat

# Custom commands via AGENTS.md or config

🔧 Function Calling / MCP

// Via API payload (function calling)
tools: [{
  type: "function",
  function: {
    name: "get_weather",
    description: "Get current weather for a location",
    parameters: { ... }
  }
}]

// Via MCP server config
mcp_servers: [{
  type: "url",
  url: "https://mcp.github.com/sse",
  name: "github-mcp"
}]

🧠 Skills (Codex CLI / IDE)

# Skills live in .agents/skills/ (repo-scoped)
.agents/skills/my-skill/
├── SKILL.md        # name, description, instructions
├── scripts/        # optional executables
└── resources/      # optional supporting files

# Install from community via skill-installer
$skill-installer install the linear skill from the .experimental folder

# Or install from a community registry
npx skills add linear

🤖 Custom GPTs (ChatGPT)

1. Go to chatgpt.com → Explore GPTs → Create
2. Use GPT Builder to define instructions, upload knowledge, add Actions
3. Publish to the GPT Store (public) or share via private link
4. Optionally add Custom Actions using an OpenAPI schema to call external APIs

📦 Plugins (Codex CLI v0.110.0+)

# Install from a marketplace
/plugin marketplace add owner/repository
/plugin install plugin-name@marketplace-name

# Codex auto-discovers installed plugins at session start
# Plugins can bundle skills, MCP servers, and app connectors

Configuration Scoping — Where You Put Things Matters

Like Claude Code, Codex configuration is directory-scoped. The directory you launch Codex from determines which instructions, skills, and settings are active.

Scope Hierarchy (highest precedence first)

How Each Config Type Resolves

Practical Rule of Thumb

• Project-specific conventions (stack, linting rules, test commands) → put in the repo’s AGENTS.md.
• Personal preferences (model, style, global instructions) → put in ~/.codex/AGENTS.md or ~/.codex/config.toml.
• Team-shared skills → commit to .agents/skills/ so everyone on the team gets them.
• Personal skills → place in ~/.agents/skills/ so they follow you across repos.

Popular Examples by Category

⚡ Slash Commands (Codex CLI)

🔧 Function Calling / MCP Tools

🧠 Skills (Codex)

🤖 Custom GPTs (ChatGPT)

Key Concepts Unique to OpenAI

AGENTS.md

The OpenAI equivalent of Claude's CLAUDE.md. A markdown file in your repo root that gives the agent persistent project context — tech stack, conventions, key commands, and guidelines. Codex reads it automatically at session start.

# Project Context
## Stack: Next.js 14, TypeScript, Supabase
## Run: npm run dev
## Test: npm test
## Conventions: functional components, no default exports

GPT Store

ChatGPT's marketplace for Custom GPTs — roughly comparable to a mix of Claude's plugin distribution and project-specific assistants. Users can discover, install, share, and in some cases monetize GPTs.

Session Forking (/fork)

Codex CLI lets you clone a conversation into a new thread with /fork (or codex fork from the shell). Useful for branching off an existing conversation to explore a different approach without losing the original.

Codex Cloud Sandboxes

Codex's cloud agent runs tasks in isolated, ephemeral sandboxes preloaded with your repository. This enables fully autonomous, asynchronous task delegation — you describe the task, Codex works independently, then returns a diff or PR. Claude's equivalent is background sub-agents.

Built-in Session Commands (Codex CLI)

Not part of the extension stack — native Codex CLI utilities

Cross-Platform Note: Skills are a Shared Standard