ChatGPT Codex: What It Is and Where It Runs

ChatGPT Codex is OpenAI’s coding agent, delivered as part of ChatGPT rather than as a separate product. As of September 30, 2026, it is one of the choices inside ChatGPT: in the desktop app you pick Codex from the ChatGPT menu, next to Chat and Work, and it reads your code, edits files, runs commands and checks its own work. Each task runs in one of two places. On your computer, Codex works in a folder you open, with the tools you already have installed. In the cloud, it works in an environment OpenAI hosts, with your GitHub repositories checked out, and keeps going while your laptop is closed. The same agent also runs in the Codex CLI and the IDE extension, and every ChatGPT plan from Free to Enterprise includes some Codex, with Free and Go limited to the desktop app.

What “Codex in ChatGPT” means today

OpenAI has folded its developer and agent products into ChatGPT. Its ChatGPT quickstart↗ (opens in a new tab) describes three experiences: Chat for quick questions, Work for research and finished documents, spreadsheets and presentations, and Codex “for software development with codebase context and developer tools.” In the desktop app you select Codex from the ChatGPT dropdown; its chats are listed separately from your ChatGPT chats, and a Quick chat icon on the New chat row handles a one-off question without opening a full task.

The names have not settled everywhere. What used to be the standalone Codex app is now this Codex view of the ChatGPT desktop app, and codex app from the terminal opens it. The cloud side is called Codex Cloud in OpenAI’s documentation, while the openai/codex README↗ (opens in a new tab) on GitHub still calls the cloud agent Codex Web and sends you to chatgpt.com/codex. Older environment pages are now labeled Codex Cloud (Legacy); OpenAI says that legacy setup still serves Code Review and the Linear and GitHub integrations. If a guide you are following uses different words, check its date.

Where Codex tasks run

On your computer

In the desktop app you start a chat, open a folder or pick a project, and Codex works on the files there. The composer shows where it will work, such as “On my computer,” and the model and reasoning control sits beneath it. Local work runs inside Codex’s sandbox and permission settings on your machine, so it can use your installed toolchain, your local services and files that are not in any repository. The CLI and the IDE extension are local in the same way.

In the cloud

OpenAI’s Codex Cloud page↗ (opens in a new tab) lays out the flow. Open ChatGPT on the web or in the desktop app, choose Work in, then Cloud, and select or create an environment. For a new one you pick GitHub repositories, connecting GitHub if prompted; Codex inspects them, installs dependencies and tools, and tests the setup, asking you for anything missing. You review the result and select Publish. After that, each task you start from the environment gets its own workspace, keeps working while your computer is asleep, and comes back with changed files and test results for you to review, follow up on, commit or open as a pull request.

• Network access is set per environment, alongside network secrets and environment variables, and you can limit the domains an environment may reach.
• On the legacy setup, OpenAI’s internet access page↗ (opens in a new tab) says agent internet access is off by default while setup scripts can still install dependencies, and it warns that turning it on raises the risk of prompt injection and code or secret exfiltration.
• Cloud needs a ChatGPT sign-in. An API key covers local work only.
• If you sign in with an email and password, you must set up multi-factor authentication before you can use Codex Cloud.

How to write tasks that suit the cloud, and what to put in an environment, is covered in Codex cloud.

Which ChatGPT plans include Codex

OpenAI says Codex and ChatGPT Work are included in the Free, Go, Plus, Pro, Business, Edu and Enterprise plans, and that the two share one usage allowance. The split, by plan name:

• Free and Go: Codex in the desktop app with GPT-6 Luna, subject to rollout. OpenAI lists no cloud features for them.
• Plus: Codex on the web, in the CLI, in the IDE extension and on iOS, plus cloud integrations such as automatic code review and Slack.
• Pro: everything in Plus with more usage, in several tiers.
• Business, Edu and Enterprise: the same surfaces inside a managed workspace, where seats and workspace roles decide which features each person can use.

So yes, you can use the Codex CLI with ChatGPT Plus: run codex, choose Sign in with ChatGPT, and the CLI draws on the same allowance as your cloud tasks. How that allowance is counted, and what to do when it runs out, is in Codex usage limits. Signing in with an API key instead is covered in Codex CLI sign-in: ChatGPT account vs API key.

When to use the CLI or the IDE extension instead

ChatGPT is the right door when you want several pieces of work in view at once, or want to hand a task to the cloud and walk away. Two other doors suit other habits:

• The Codex CLI, if you live in a terminal, work over SSH, or want Codex in a script or CI job through codex exec. It can also start and list cloud tasks with codex cloud.
• The IDE extension, if you want Codex beside the file you are editing, with open files and selections as context and diffs you accept hunk by hunk. It can hand longer work to the cloud too.

All of them share one sign-in and the same AGENTS.md instructions, and the CLI and extension share config.toml. The full surface-by-surface comparison is in Codex app vs CLI vs IDE extension.

codex login status    # which sign-in is active
codex app             # open Codex in the ChatGPT desktop app
codex cloud           # browse cloud tasks or start one from the terminal

ChatGPT agent vs Codex, and the ChatGPT app vs the Codex app

ChatGPT agent mode is no longer available; its job moved to Work, which takes an outcome and returns a document, spreadsheet, presentation or report. Codex is the one for code: it works against a repository, runs your tests and ends in a diff or a pull request. If the deliverable is a file for people to read, use Work; if it is a change to a codebase, use Codex. The history is in ChatGPT agent mode and what replaced it.

There is no longer a separate Codex desktop app to choose against the ChatGPT app. OpenAI’s desktop app page↗ (opens in a new tab) says you install ChatGPT for macOS, Windows or Linux and then choose ChatGPT or Codex inside it. The Linux build is in preview.

When the Codex connection is not working

Most “Codex can’t see my repository” reports come down to one of a few settings. Work through them in order:

1. Check the sign-in. Cloud tasks need Sign in with ChatGPT; a session signed in with an API key cannot reach them.
2. Check GitHub. The environment setup asks you to connect GitHub and choose repositories; a repository you did not select is invisible to Codex.
3. Check the environment was published. A draft environment cannot start tasks.
4. Check multi-factor authentication if you sign in with an email and password.
5. In a Business or Enterprise workspace, ask an admin whether your role allows Codex Cloud.

For ChatGPT’s other connectors, the general checklist is in ChatGPT connectors not working.

Where the task list lives

Codex keeps a history of chats per surface: desktop threads, CLI sessions and cloud tasks each have their own. None of them is a list of what the team still has to do. fenbs is that list, a board your teammates and your AI assistants share, with lanes To Do, Next Up, In Progress and Completed. Each task holds a note with the problem, a plan with the steps, and a test status with notes when it is done. The local Codex surfaces read MCP servers from config.toml, so adding https://fenbs.ai/api/mcp once lets the CLI and the extension read a task with fenbs_get_item, move it along and comment on it, and History records each change under the assistant’s name. For a cloud task, paste the task’s note and plan into the prompt so the cloud run starts from the same definition of done. fenbs has no due dates, sprints or settable assignee; the lanes and a 1 to 10 priority do the ordering.