#Introduction to Codex

#What is Codex?

Codex is OpenAI's AI coding agent—not just a code-completion model. In your local workspace it can:

• Read codebases and documentation
• Edit files and run commands
• Verify results (tests, lint, builds)
• Iterate until the task is done or needs your approval

Codex has evolved through two major shifts:

This tutorial centers on Codex CLI (comparable to Claude Code) and covers how App, IDE extensions, and Cloud fit together.

#Product matrix: four surfaces, one agent stack

                    ┌─────────────────────────────────┐
                    │     OpenAI Codex (same stack)   │
                    └─────────────────────────────────┘
           ┌────────────┬────────────┬────────────┬────────────┐
           │  Codex CLI │  Codex App │ IDE ext.   │ Codex Cloud│
           │  Terminal  │ Desktop GUI│ VS Code…   │ Browser/PR │
           └────────────┴────────────┴────────────┴────────────┘
                    local workspace          remote env + GitHub

#1. Codex CLI (tutorial focus)

• Open source: github.com/openai/codex
• Rust implementation: fast startup, low overhead
• Interactive TUI, codex exec for non-interactive runs, sandbox and approval policies
• Best for: daily development, script automation, CI assistance

#2. Codex App

• macOS / Windows desktop app
• Pick a project folder; runs locally in Local mode
• Best for: developers who prefer a GUI and visual diffs

#3. IDE extensions

• VS Code, Cursor, Windsurf, and others
• Sidebar agent mode with read/write and run permissions by default (configurable)
• Best for: inline edits while coding

#4. Codex Cloud

• Visit chatgpt.com/codex
• Connect GitHub repos; tasks run in cloud environments and produce PRs
• @codex in GitHub PR comments delegates work
• Best for: async long tasks and cross-timezone collaboration

#Core capabilities

#Code understanding and generation

• Generate or modify code from natural language
• Cross-file refactors, tests, and documentation
• Image input (screenshots, designs) and image generation/editing (CLI capabilities)

#Agent toolchain

Codex works through built-in tools and extension protocols:

• File read/write and patch application
• Shell commands (within sandbox policy)
• Code search (glob, grep, etc.)
• Web search (cached / live modes)
• MCP: connect Linear, GitHub, docs services, and other external systems

#Customization stack

OpenAI splits “make the agent work like our team” into five layers that complement each other:

Understanding these five layers is the step from “using Codex” to “using it well.” Later chapters cover each layer.

#Good fit vs. caution

Good fit:

• Feature work, bug fixes, and tests in familiar repos
• Bulk refactors, migrations, and doc sync
• Local code review (dedicated review agent)
• Turning repeat flows into Skills or codex exec scripts

Use caution:

• Production secrets or unisolated sensitive environments (always use sandbox + approvals)
• Repos with no tests or lint (“bare” repos where the agent may loop on mistakes)
• Fully unattended auto-merge (prefer PR + human review)

#Comparison with other tools (brief)

Choosing: If you already have ChatGPT Plus/Pro and want local agent + Cloud PRs + open AGENTS.md, Codex is a full stack. Deep Anthropic usage favors Claude Code; IDE-only completion favors Copilot.

#Billing and accounts

• ChatGPT Plus / Pro / Business / Edu / Enterprise plans typically include Codex usage
• OpenAI API key login is also supported (some features may be limited)
• CLI ChatGPT login can configure the API key automatically

See OpenAI Codex docs for current quotas.

#Next steps

• Installation: install the CLI and authenticate
• Quick Start: complete a small verifiable task
• Architecture & Runtime: how the agent plans, executes, and respects the sandbox