Getting Started with Codex CLI
From the course OpenAI Codex CLI
OpenAI Codex CLI is an open-source, lightweight terminal agent for AI-assisted coding. It connects to OpenAI's models to provide code generation, editing, and command execution directly from your terminal.
Installation
npm install -g @openai/codex
export OPENAI_API_KEY=your-key-here
codex
Key Features
- Open-source — MIT licensed, community-driven development
- Lightweight — Minimal dependencies, fast startup
- Sandboxed — Network-disabled, directory-restricted by default
- Multi-model — Works with GPT-4o, o3, o4-mini, and other OpenAI models
- Cross-platform — macOS, Linux, and WSL on Windows
Approval Modes
| Mode | File Edits | Commands | Best For |
|---|---|---|---|
| Suggest | Requires approval | Requires approval | Learning, sensitive code |
| Auto-edit | Automatic | Requires approval | Routine development |
| Full-auto | Automatic | Automatic | Trusted automation |
Basic Usage
# Interactive mode
codex
# One-shot command
codex "add input validation to the signup form"
# With specific model
codex --model o3 "refactor this module"
# Quiet mode (less output)
codex -q "fix the failing test"
Safety Model
Codex CLI applies sandboxing by default:
- Network access disabled in sandbox
- File access restricted to current directory and children
- Commands run in isolated environment
- Full-auto mode applies additional restrictions
- Review all changes before committing
Codex CLI in a Real CI Pipeline
One practical pattern: run Codex CLI in GitHub Actions to auto-fix lint errors on every pull request. The workflow installs the CLI, runs your linter, pipes failures to Codex, and commits the fixes back to the branch — all without human intervention for low-risk style violations.
# .github/workflows/auto-lint-fix.yml (simplified)
- name: Run linter and auto-fix
run: |
npm run lint 2>&1 | codex -q --model o4-mini "Fix all ESLint errors shown in stdin. Only modify formatting, do not change logic." --allowedTools edit
Keep the model at o4-mini for this task — it is fast, cheap, and more than capable of formatting fixes. Reserve o3 for architectural decisions.
Best Practice — Commit Codex Changes Separately
When Codex makes edits in auto-edit mode, stage and commit them as a discrete "chore: AI-assisted refactor" commit rather than mixing them with your own manual changes. This makes the diff reviewable, keeps blame history clean, and gives you a clear rollback point if the AI introduced a subtle regression.
Common Pitfall — Running in Full-Auto on Shared Branches
Full-auto mode executes shell commands without prompting. Never run it on a branch that other people are actively working on — a runaway command could delete build artefacts or overwrite files mid-pipeline. Use full-auto only on isolated feature branches or disposable worktrees.
Key Takeaways
- •Codex CLI is open-source (MIT) and lightweight for terminal-based AI coding
- •Three approval modes: suggest, auto-edit, and full-auto
- •Sandboxed by default for safety — no network, restricted directory access
- •Supports multiple OpenAI models including GPT-4o and o3
Progress is kept in this browser only.