OpenAI Codex: AI Coding Agent Guide
What is OpenAI Codex?
Codex is OpenAI's AI coding agent that works in your terminal, IDE, and the cloud. It can write features, fix bugs, answer questions about your codebase, and propose pull requests—all running in parallel sandbox environments preloaded with your repository.
Codex is powered by GPT-5.3-Codex (released February 2026), the most capable agentic coding model to date. It combines frontier coding performance with strong reasoning capabilities and runs 25% faster than previous versions.
Where Codex Runs
| Platform | How to Access |
|---|---|
| Terminal (CLI) | npm i -g @openai/codex or brew install --cask codex |
| VS Code / Cursor / Windsurf | Install the Codex IDE extension |
| Cloud (Codex App) | chatgpt.com/codex |
| Slack | @codex mentions for questions and tasks |
| GitHub | @codex on PRs and Issues |
What Makes Codex Unique
- Parallel execution: Run multiple tasks simultaneously in isolated cloud sandboxes
- Skills integration: Native support for the Agent Skills standard (
SKILL.md) - Automations: Unprompted background work—issue triage, CI/CD, alert monitoring
- @codex mentions: Tag Codex on GitHub PRs/Issues or in Slack to assign tasks
- Agent Teams mode: Multiple Codex instances working together on complex projects
Is Codex Free?
Codex is included with ChatGPT Plus ($20/month), Pro, Business, Edu, and Enterprise plans. There's no separate subscription for Codex.
The Codex CLI is open-source and free to install:
npm i -g @openai/codex
# or
brew install --cask codex
You can also point the CLI at any model provider (not just OpenAI) that supports the Chat Completions API.
Codex Models
| Model | Best For | Availability |
|---|---|---|
| GPT-5.3-Codex | Complex coding, reasoning, large refactors | ChatGPT subscriptions, Codex App/CLI/IDE |
| GPT-5-Codex-Mini | Faster, cost-effective (4x more usage) | ChatGPT subscriptions |
| Custom models | Point Codex CLI at any compatible API | API key required |
How Codex Skills Work
Codex natively supports the Agent Skills standard. Skills provide:
- Task-specific workflows: Code review, testing, deployment, documentation
- Domain expertise: SEO, frontend design, database optimization
- Team conventions: Your coding standards, review processes, testing requirements
Skill Invocation Methods
- Explicit: Type
$skill-namein your prompt - Implicit: Describe a task—Codex auto-selects the matching skill
- List skills: Run
/skillsto see all available skills
Codex uses progressive disclosure: it loads only skill metadata at startup (~100 tokens each), then fetches full instructions only when a skill is invoked.
How to Install Skills on Codex
Install the /learn Command
The /learn command connects Codex to 40,000+ skills from agentskill.sh:
git clone https://github.com/agentskill-sh/learn.git ~/.codex/skills/learn
Then in any Codex session:
/learn seo # Search for skills
/learn @anthropic/seo-content-optimizer # Install specific skill
/learn trending # See popular skills
Manual Installation
- Browse skills on agentskill.sh
- Download the skill ZIP
- Extract to
~/.codex/skills/[skill-name]/
Codex auto-discovers skills on startup.
Using AGENTS.md
For project-specific customization, add an AGENTS.md file to your repository. This tells Codex:
- How to navigate your codebase
- Which commands to run for testing
- Your project's coding standards
- Preferred patterns and conventions
Popular Codex Skills
| Skill | Description | Installs |
|---|---|---|
| code-review | Systematic review with security & performance checks | 15K+ |
| test-runner | Run tests, analyze failures, coverage reports | 12K+ |
| frontend-design | Production-grade UI components | 8K+ |
| seo-optimizer | SEO, keyword research, content optimization | 10K+ |
| programmatic-seo | SEO pages at scale | 6K+ |
Codex vs Claude Code
Both are powerful coding agents with different strengths:
| Feature | Codex | Claude Code |
|---|---|---|
| Maker | OpenAI | Anthropic |
| Model | GPT-5.3-Codex | Claude Opus 4.5 |
| Cloud execution | Yes (sandboxed) | No (local only) |
| GitHub integration | @codex mentions | Git CLI |
| Slack integration | Yes | No |
| Skills format | SKILL.md | SKILL.md (same) |
| Pricing | ChatGPT Plus ($20/mo) | Claude Pro ($20/mo) |
Both use the same Agent Skills standard, so skills are fully interoperable.
FAQ
Is Codex free?
The Codex CLI is free and open-source. Full Codex functionality (cloud tasks, parallel execution, GitHub/Slack integration) requires a ChatGPT Plus subscription ($20/month) or higher. You can also use the CLI with any compatible API provider.
What's the difference between Codex and ChatGPT?
ChatGPT is a conversational AI. Codex is an autonomous coding agent built on top of GPT-5.3-Codex—a model specifically optimized for software engineering. Codex can:
- Execute code in sandboxed environments
- Work on multiple tasks in parallel
- Integrate with GitHub and Slack
- Remember project context across sessions
Can I use Codex with Claude or other models?
Yes. The Codex CLI supports any model provider that implements the Chat Completions or Responses API. You can point Codex at Anthropic Claude, local models via Ollama, or any compatible endpoint.
How do Skills differ from Automations?
- Skills: Task-specific capabilities you invoke (e.g., "Review this code"). They provide structured workflows for specific jobs.
- Automations: Unprompted background work Codex performs autonomously—issue triage, CI/CD monitoring, alert handling.
Skills are reactive (you ask). Automations are proactive (Codex acts on its own).
Can Codex use Claude Code skills?
Yes. Both Codex and Claude Code use the same Agent Skills standard (SKILL.md format). Skills are fully interoperable—write once, use everywhere.
How do I customize Codex for my project?
Add an AGENTS.md file to your repository root. This is like a README for Codex—it tells the agent how to navigate your codebase, which commands to run, and your project's conventions. You can also install skills that match your tech stack.
What is Agent Teams mode?
Agent Teams (research preview in Opus 4.6) allows multiple Codex instances to work in parallel and coordinate autonomously. One agent might research while another writes code and a third runs tests—all synchronized.
Is Codex secure?
GPT-5.3-Codex is treated as a high-security capability under OpenAI's Preparedness Framework. Cloud tasks run in isolated sandboxes. For sensitive projects, you can run the CLI locally with your own API keys.
Sources: