Codex CLI

Codex CLI is OpenAI's open-source coding agent that runs entirely in your terminal. You point it at a repository, describe a task in plain language, and it reads files, edits them on disk, and runs shell commands to get the job done — all inside an OS-level sandbox that blocks network access by default and, once you trust the folder, scopes writes to your workspace. It is written in Rust and ships as a binary installable via npm, Homebrew, or a one-line shell installer.

It is aimed at developers who live in the terminal and want an agent backed by OpenAI's frontier models without leaving the shell. You can authenticate with a ChatGPT plan (Free, Go, Plus, Pro, Business, Edu, or Enterprise) or an OPENAI_API_KEY, and the same binary works on macOS, Linux, and Windows (natively or via WSL2). Beyond the terminal, Codex ships an IDE extension and the Codex cloud agent, and codex app opens the desktop experience, which has been part of the ChatGPT desktop app on macOS and Windows since July 2026.

Highlights

• Two-layer security model — sandbox modes (read-only, workspace-write, danger-full-access, via --sandbox) control what the agent can technically do; approval policies (on-request, never, or a granular per-category policy) control when it must stop and ask before acting. The untrusted approval value was retired in August 2026 and now prevents Codex from starting, so remove it from older configs; the per-project trust_level = "untrusted" setting still works.
• Sandboxed by default — outbound network is blocked by default, and in a trusted project the default workspace-write mode limits writes to the active workspace (a folder you haven't trusted yet can start in read-only), so edits stay local until you explicitly widen the boundary.
• Model switching — use /model to choose among the models available to your account and adjust reasoning effort per task, or pass --oss to run local open-source models through Ollama or LM Studio.
• MCP support — connect external tools through STDIO or Streamable HTTP servers configured in config.toml.
• Non-interactive codex exec — run Codex headlessly in scripts and CI, piping the final result to stdout.
• Session resume and image input — pick up past transcripts with codex resume, and attach screenshots or design specs as context.

In an AI-assisted workflow

Codex CLI fits where you already run Git and your build. A typical loop is to start it in a trusted repo with the default workspace-write sandbox mode and on-request approval policy, let it draft edits, and approve anything that reaches outside the workspace or touches the network. It reads AGENTS.md files for project-specific context, so you can encode conventions and commands once and have them apply on every run. Repeatable procedures can move into Codex skills instead of growing the always-on instruction file.

npm install -g @openai/codex
cd your-project
codex "Add a retry with backoff to the API client and a test for it"

Good to know

Codex CLI is free and open source under the Apache-2.0 license, available on macOS and Linux natively and on Windows (natively via PowerShell or under WSL2). Model usage draws on your ChatGPT plan's included Codex allowance (even the Free and Go plans include one) or is billed per token with an API key. The danger-full-access sandbox mode removes network and filesystem guardrails — use it only on repositories and tasks you fully trust.