OpenAI Codex

OpenAI Codex is OpenAI's coding agent, and it meets developers in several places: as a command-line program in the terminal, as an extension for editors such as Visual Studio Code, Cursor, and Windsurf, as a desktop app, and as a web-based agent inside ChatGPT. According to the provider, Codex takes tasks described in natural language, reads and edits files in a project, runs commands, and hands back changes you can review and accept. The command-line version is open source under the Apache 2.0 license; its public GitHub repository counts well over a hundred thousand stars, making it one of the most closely followed AI coding projects, and it is being developed on a daily basis.

The variants at a glance

The Codex CLI is the local variant: it runs on your own machine, works directly in the chosen project folder, and suits anyone whose workflow already lives in the terminal. The IDE extension brings the same agent into the editor so that suggestions and changes appear next to your own code. The desktop app bundles several tasks, an integrated terminal, and Git worktrees in one place. The web variant at chatgpt.com/codex runs tasks in a cloud environment, which fits longer-running or parallel jobs without tying up your own computer.

Installation and sign-in

According to the repository, the CLI can be installed with an install script, via npm, via Homebrew, or as a prebuilt binary from the GitHub releases page; Windows has a PowerShell route. After installing, you start the agent with the codex command. There are two ways to sign in: either with your ChatGPT account, using the quotas included in your plan, or with your own API key, billing usage at the regular API terms. Per the documentation, the CLI, the SDK, and the IDE extension work with an API key, whereas the cloud features do not.

Security, sandbox, and approvals

A central trait of Codex is how it handles permissions. Commands run in a sandbox, and approval modes decide how independently the agent may act, from confirming every single action to largely automatic runs inside the sandbox. This lets you tune the balance between speed and control to the project and your level of trust. For organizations, the provider also mentions automated reviews and role-based profiles.

Extensibility: AGENTS.md, MCP, and skills

Project-specific behavior is described in a file called AGENTS.md, where build commands, conventions, or off-limits areas of the repository are recorded. Through the Model Context Protocol, Codex connects external tools and data sources, and skills let you bundle recurring procedures as reusable building blocks. Anyone already using MCP servers from the Skill Road catalog can therefore use them with Codex as well.

Pricing and plans

According to the provider, Codex is included in all ChatGPT plans, from the free entry level through the individual plans up to Business, Enterprise, and Education; quotas differ by plan, and some plans apply rolling time windows. If you would rather pay by usage, you use an API key. Current terms are on the provider's official pricing page and tend to change often, which is why Skill Road deliberately names no amounts here.

Who Codex fits

Codex suits developers who prefer an agent in the terminal, value sandboxing and approvals, and already work in the OpenAI ecosystem. Teams benefit from the fact that the same configuration via AGENTS.md applies across CLI, IDE, and cloud. If you want a fully model-independent tool, the catalog offers alternatives such as Cline; if you mainly need chat and research, ChatGPT itself is often the better fit. To get started, it is enough to install the CLI, sign in, and give it a small, clearly scoped task in a test project.