Cloud Parts

OpenAI Codex

OpenAI’s coding agent: an open-source CLI, written in Rust, that works in your local repository inside an OS sandbox, plus an IDE extension and Codex cloud for tasks that run on OpenAI’s machines.

What it is
AI coding agents (terminal agents)
Type
Commercial product
Made by
OpenAI
License
Apache 2.0 (CLI)
Facts checked
9 October 2026
Website Source code

What it is

Codex CLI is a single native binary (codex) that you run in a project directory and sign into with a ChatGPT plan or an API key. It reads files, edits them and runs commands in a loop, and two settings bound what it does. The sandbox mode (read-only, workspace-write or danger-full-access) is enforced by the operating system: Seatbelt via sandbox-exec on macOS, bubblewrap plus seccomp on Linux. The approval policy decides when it stops to ask. In version-controlled folders it starts in the Auto preset, which edits and runs commands in the workspace freely but asks before writing outside it or using the network; network access is off by default and .git stays read-only even inside the workspace.

Before any work it reads AGENTS.md files: a global one in ~/.codex, then one per directory from the repository root down to where you started, concatenated so the closest file comes last, up to 32 KiB by default. codex exec runs one task non-interactively for scripts and CI, codex mcp adds MCP servers, and codex review reviews a diff without touching the working tree.

The same agent runs elsewhere. The IDE extension works in VS Code, Cursor and Windsurf, with separate integrations for JetBrains IDEs and Xcode. Codex cloud runs each task in its own workspace in an OpenAI-hosted environment built from your GitHub repositories, keeps working while your computer sleeps, and ends with a diff you can turn into a pull request; codex cloud submits and applies those tasks from the terminal. The name is older than the agent: in 2021 “OpenAI Codex” was the model behind GitHub Copilot’s first preview.

Why teams pick it

  • The CLI is open source under Apache 2.0 and ships as a native Rust binary, with no Node.js runtime to install
  • Sandboxed by default, with the OS enforcing workspace-only writes and no network unless you allow it
  • One agent across terminal, IDE, cloud tasks and GitHub code review
  • AGENTS.md, which OpenAI introduced, is now a vendor-neutral format read by many other agents

Watch out for

  • Built around OpenAI models and accounts; OpenAI recommends signing in with a ChatGPT plan
  • The Linux sandbox can fail inside Docker when the container blocks namespaces, setuid bwrap or seccomp; the docs suggest a Dev Container as the outer boundary instead
  • Codex cloud needs an environment that mirrors yours (repositories, setup, secrets, network rules) and works from GitHub repositories
  • Easy to confuse with the 2021 Codex model of the same name, and with the separate Codex Security product

Alternatives

What all terminal agents have in common →
Claude CodeCommercialAnthropic’s coding agent: a terminal program, also available in IDEs, a desktop app and the cloud, that drives Claude models through a loop of reading, running and editing.Gemini CLICommercialGoogle’s open-source terminal agent for Gemini models; since 18 June 2026 it serves only paid API keys and business licences, as Google moved free and consumer users to Antigravity CLI.OpenCodeOpen sourceAn open-source (MIT) coding agent that works with 75+ model providers and local models, built as a local server with a terminal interface as one of its clients.+1 Kilo Code

How it fits with other technologies

Choose the sandbox, then run interactively or from a script · bash

curl -fsSL https://chatgpt.com/codex/install.sh | sh   # or: npm install -g @openai/codex

codex                                   # interactive, Auto preset in a git repo
codex --sandbox read-only --ask-for-approval on-request   # look, don’t touch

# one task, no session: for scripts and CI
codex exec "run the tests and fix the first failure"

codex cloud                             # browse Codex cloud tasks, apply results locally

See it in a real stack

No architecture uses OpenAI Codex yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2021–2025

OpenAI launched Codex CLI in April 2025 as a lightweight, open-source coding agent for the terminal, running models such as o3 and o4-mini. A month later, on 16 May 2025, it announced Codex, a cloud agent in research preview: each task ran in its own cloud sandbox preloaded with the repository, typically took 1 to 30 minutes, ran with internet access disabled, and could be guided by AGENTS.md files in the repository. It was powered by codex-1, a version of o3 tuned for software engineering.

The first CLI was written in TypeScript. On 30 May 2025 OpenAI announced it was rewriting it in Rust, describing the core of the project as “an agentic harness, aka calling the model in a loop”, and naming the reasons: a zero-dependency install instead of requiring Node.js 22, native security bindings (it already shipped Rust for Linux sandboxing) and performance. In September 2025 GPT-5-Codex arrived along with an IDE extension for VS Code and its forks, a rebuilt CLI and three simple approval levels. AGENTS.md, released by OpenAI in August 2025, moved to the Linux Foundation’s Agentic AI Foundation that December.

Why it’s built this way

Open-source the CLI
The terminal agent runs on developers’ machines with their files and credentials. Publishing it under Apache 2.0 let users inspect what it does and contribute; OpenAI credited community feedback for the September 2025 rebuild.
Rewrite in Rust
Requiring Node.js 22 was a blocker for some users, and the Linux sandbox was already Rust. A native binary removed the runtime and the garbage collector from the loop.
Sandbox first, ask second
The cloud agent started with internet access off during tasks, and the CLI enforces workspace-only writes at the OS level, so approvals are needed only for leaving the sandbox rather than for every command.
Instructions in a plain Markdown file
AGENTS.md gives agents build, test and style instructions without cluttering the README, under a name that isn’t tied to one vendor.

Timeline

  1. 2021

    The original Codex model powers GitHub Copilot’s technical preview

  2. 2025

    Codex CLI launches as an open-source terminal agent (April)

  3. 2025

    Codex cloud research preview, powered by codex-1 (May)

  4. 2025

    Rust rewrite of the CLI announced (May)

  5. 2025

    GPT-5-Codex, the IDE extension and a rebuilt CLI (September)

  6. 2025

    AGENTS.md moves to the Agentic AI Foundation (December)

Sources7 · facts checked 9 October 2026

Improve this page on GitHub