The Codex app: what it is and how to get it

What the Codex app is in August 2026

OpenAI released a standalone Codex desktop app for macOS on 2 February 2026 and for Windows in early March 2026. In July 2026 it was merged into the ChatGPT desktop app, where Codex became one of three modes alongside Chat and Work. Existing installs updated in place and renamed themselves, which is why a lot of people have the app and cannot find it.

The official documentation reflects the merge: the page at learn.chatgpt.com/codex/app is titled ChatGPT desktop app, and the surface list on the Codex overview reads ChatGPT desktop app, Remote, ChatGPT on the web, Codex CLI, Codex IDE extension, Codex cloud. The brand survives in the marketing URL: the Codex CLI README still points at a "Codex App page" at chatgpt.com/codex?app-landing-page=true.

Getting it

codex app     # opens the ChatGPT desktop app

What the app does that the CLI does not

The app is not a graphical wrapper around codex. It exists because a terminal is a poor place to supervise five agents at once, and everything it adds is about that problem.

Parallel chats in managed worktrees

Several chats can run against one project at once because each can take its own git worktree, so an agent refactoring a module cannot collide with one writing tests. The app manages them: worktrees are created under $CODEX_HOME/worktrees by default, the root is configurable in Settings, Worktrees, and any setup script you configure for the local environment runs automatically when a new one is created.

Review in the app instead of a scrollback

The diff pane shows the current checkout, takes inline comments for Codex to address, and stages or reverts individual hunks or whole files. You can commit, push a branch and open a pull request without leaving it. Typing /review in the composer asks Codex to review either against a base branch or just the uncommitted changes. The CLI equivalent is reading a transcript and running git diff in each worktree yourself, which is fine for one task and unmanageable for six.

Automations

Scheduled and repeatable work: a task that runs on a cadence rather than when you type. In a git repository those scheduled runs get dedicated background worktrees so they cannot disturb what you are doing; in a project with no version control they run in the project directory itself, which is worth knowing before you schedule anything destructive. The CLI equivalent is codex exec under cron, which is more flexible and considerably more setup.

Remote control from your phone

Settings, Connections, "Control this Mac or PC" produces a QR code. Scan it with the ChatGPT mobile app, sign into the same account, approve the connection, and the Remote tab on the phone can watch running tasks, approve or deny commands, read diffs and send new instructions. The work still happens on your machine, so the machine has to stay awake and online.

The app and the CLI are one product

This is the part worth internalising, because it removes most of the anxiety about picking wrong. Both front ends work out of the same Codex home directory, which is ~/.codex unless CODEX_HOME says otherwise. The app's own worktree root defaults to $CODEX_HOME/worktrees, which is the giveaway.

~/.codex/config.toml    # model, sandbox_mode, approval_policy, MCP servers
~/.codex/auth.json      # CLI credentials
~/.codex/sessions/      # JSONL transcripts
~/.codex/worktrees/     # the app's managed worktrees

codex login status      # who am I signed in as
codex app               # open the desktop app

So an AGENTS.md you wrote for the CLI applies in the app. An MCP server configured in config.toml is available in both. And usage from both counts against the same ChatGPT plan limits, which is the single most common billing surprise: running the app and the CLI at once does not give you two allowances.

When the app is the wrong choice

Questions people ask

Keep reading