What Is OpenAI Codex
OpenAI Codex is an AI coding agent OpenAI introduced in 2025. Its biggest trait is "throw a task and it's done while you sleep" async execution. There are CLI and Cloud forms.
2 Forms
1. Codex CLI
A CLI running in the local terminal. Operation feel similar to Claude Code.
2. Codex Cloud (Agent Mode)
Runs in OpenAI's cloud. Links with a GitHub repo and, triggered by Issues or PRs, automatically writes code and submits a PR.
Pricing
- Included from ChatGPT Plus (USD 20/mo) up
- Pro (USD 200/mo): greatly increased usage
- Via API: usage billing, separate
Codex CLI Install
npm install -g @openai/codex codex login # Log into OpenAI in the browser cd your-project codex
Usage (CLI)
Interactive Mode
$ codex > What's this project's structure? > Rewrite the auth module to JWT auth > Add tests too
One-Line Task
$ codex "make all TypeScript files strict-mode compliant"
Codex Cloud Setup
1. GitHub Linkage
- chat.openai.com → Settings → Codex → GitHub linkage
- Select target repos
- Access: Read / Write, Issue / PR operations
2. Submit a Task
- Select "Codex" mode in ChatGPT
- Instruct "fix bug X in this repo and create a PR"
- Codex branches in the cloud, writes code, makes a PR
- Track progress in the ChatGPT screen
Strengths
- Async: a PR is complete while you sleep
- Parallel: submit multiple tasks at once
- GitHub integration: Issue → PR end to end
- Reliable exec environment: runs tests in an isolated cloud environment
Weaknesses
- Pricing: long tasks require the Pro plan
- Supervision: runs in the cloud, real-time intervention is hard
- Security: confidential repos with care
- Less adopted vs Claude Code: ecosystem still smaller
Customize with AGENTS.md
Place AGENTS.md in the project root and Codex always follows it:
# AGENTS.md ## Project rules - TypeScript strict - All changes must include tests - Run npm run lint before finalizing - Commit messages: Conventional Commits format
Typical Use Cases
- "Fix this bug group" in the morning → PR complete by noon
- "Test coverage to 80%" → achieved in a few hours
- "Update dependencies to latest" → auto up to test confirmation
- "Address all security warnings" → Issue to PR
Selective Use vs Claude Code
| Codex Cloud | Claude Code | |
|---|---|---|
| Where it runs | Cloud | Local CLI |
| Async | ◎ | ○ |
| Supervision | Hard | Easy |
| Security | Cloud-sent | Local-centric |
| Pricing | Plus ~ Pro | Pro ~ Max |
Mastery Tips
- Issue-driven: write details in a GitHub Issue and pass to Codex
- Approval flow: auto PRs always human-reviewed
- Mandatory tests: guarantee AI-code quality in CI
- From routine tasks: dependency updates, test additions, lint fixes
Next Step
The complete guides for the 4 main tools (Claude Code / Cursor / Copilot / Codex) are now in place. Next, to Skills and AI pair-programming know-how.
