OpenAI Codex CLI
OpenAI's open source coding agent that runs in your terminal with your own API key.
Description
OpenAI Codex CLI is OpenAI's coding agent that runs locally in the terminal, open source and actively maintained (v0.120.0 as of April 2026). It's written mainly in Rust, installs via 'npm install -g @openai/codex' or 'brew install --cask codex', and supports macOS, Linux, and Windows via WSL. It lets you chat with the agent to read the repo, edit files, run commands in a configurable sandbox, and validate changes, with support for AGENTS.md, skills in .codex/skills, MCP servers, and multi-step plans. It authenticates with a ChatGPT Plus/Pro/Business/Enterprise account or with your own OpenAI API key (BYOK), making it free as software (you only pay API usage or use your ChatGPT plan quota). It's OpenAI's official alternative to Claude Code, Cursor CLI, or Aider.
Preview
Detailed Evaluation
Key strengths
OpenAI's official open source agent
Code on GitHub, open to contributions, with frequent releases (v0.120.0 in April 2026) and actively maintained by the OpenAI team.
Two authentication models
You can use your ChatGPT Plus/Pro/Business/Enterprise account without paying for the API separately, or bring your own API key for isolated billing.
Configurable sandbox
Control what the agent can read, write, and execute, with adjustable approval modes to balance speed and safety.
AGENTS.md and local skills
You can instruct the agent at the repo level with AGENTS.md and define reusable skills in .codex/skills, just like other modern agents.
Native MCP
Model Context Protocol is supported out of the box, letting you plug tools like Supabase, GitHub, Linear, or Notion without friction.
Limitations to consider
Less mature than Claude Code
Claude Code has been in the terminal longer and has a slightly more developed ecosystem of skills and hooks than Codex CLI.
Non-trivial API cost
If you use BYOK with top models, tokens add up fast; without a ChatGPT quota included, long sessions can get expensive.
CLI-only
There's no graphical UI or native VS Code integration; if you want an IDE, look at Cursor or Windsurf.
Coexists with OpenAI Agents SDK
OpenAI's strategy mixes Agents SDK (framework) and Codex CLI (product), which can create confusion about which to use when.
Standout Feature
Being the only OpenAI-backed CLI agent, open source and with dual authentication (ChatGPT account or BYOK), which makes it the natural choice for anyone who wants an official agent with no black box and aligned with OpenAI's latest models.
Comparison with Alternatives
Versus Claude Code it's less polished on skills but aligned with OpenAI; versus Cursor it offers a pure open source CLI experience, no IDE; versus Aider it has better MCP integration, official support, and much more frequent releases.
Ideal User
Advanced developers who work in the terminal, want an official open source OpenAI agent, care about controlling the sandbox and skills, and already pay for a ChatGPT plan they can draw quota from. Also a fit for teams that prefer BYOK to separate spend by project.
Learning Curve
Install and first commands are simple, but mastering AGENTS.md, skills, MCP servers, and sandbox policies takes reading docs and experimenting. Developers with prior Claude Code or Aider experience will adapt in hours.
Best For
- Developers who live in the terminal and want an official OpenAI agent
- Teams already paying for ChatGPT Plus/Business who want to leverage their quota
- Projects that need to control the execution sandbox and write policies
- Integrations with MCP servers and local skill definitions
- Users who prefer an open source tool over a closed black box
Not Ideal For
- People who prefer a graphical IDE over a CLI
- Teams that need enterprise support with an SLA on the tool
- Users who don't want to manage any file-based configuration