Tool brief

OpenAI Codex Guide: App, CLI, IDE and Cloud

Learn what OpenAI Codex does across app, terminal, editor, cloud and GitHub review. Pilot repository access, tests, usage and the handoff.

Editorial update

Decision in brief

Use Codex in the CLI or IDE for a patch you want to inspect as it develops. Use the app when you want a dedicated place to manage work. Delegate work to Codex Cloud only when the repository, test setup and data access can be reproduced there. Add GitHub code review as a separate check on an eligible repository after testing feedback quality.

Choose a work surface: CLI or IDE — Inspect edits. App — Manage work. Cloud — Repeat setup.
Delegate only when the repository, checks and necessary access can be reproduced.

Choose the surface by where work happens

CLI and IDE workflows start with the checked-out project. Codex Cloud runs delegated tasks in a separate environment, which requires provisioning dependencies and test setup independent of your laptop. The app is another entry point with its own controls. To review a PR on an enabled GitHub repository, request @codex review or configure automatic review. Keep changes, review findings and approvals distinct when scoring the pilot.

Work you need to doStart hereWhat to verify
Repair a failing regression with a developer presentCLI or IDEUse the same worktree and test command and inspect the patch alongside the failure.
Keep multiple assignments organizedCodex appConfirm task context, ownership and how you will inspect each diff.
Delegate a task while you do other workCodex CloudRepository access, reproducible setup, network controls and ability to run the same suite.
Get a PR second opinionGitHub review integrationCorrect pull request, scope of rules, usefulness and review ownership.

Price the workflow you actually use

As checked November 16, 2026, OpenAI publishes Codex access and included usage by ChatGPT account plan. Usage varies by task and model, so copying another person’s prompt count gives a weak capacity estimate. Capture the account’s reported usage before and after equivalent runs. Keep included ChatGPT usage, any purchased credits and separately metered API-key usage in separate rows in your cost sheet.

What Hacker News cases make measurable

Simon Willison used Codex CLI in a JavaScript port of JustHTML and attributed the feasibility to an existing implementation-independent HTML parsing suite. That gives a clear pattern for delegated tasks. Pick behavioral tests you can run unchanged on both implementations and retain any manual inspection of gaps.

In a different thread about CLI background processes, icantevenhold reported considerable disk use by old versions on one machine. Check version management and disk use on your own managed devices. This report describes one installation and does not establish a current storage requirement.

Pilot local versus delegated work

Pick one small task in a disposable approved repository and define the expected unit-test failure before turning it over to Codex. Start each attempt from the same commit.

  1. Run a manual baseline and capture the test output and elapsed time. Set an upper limit for active review time and billable usage before starting.
  2. Try the CLI or IDE first and record each requested approval, changed file and test run. If you have a viable cloud environment, repeat from another fresh branch there with the same checks and a minimal repository grant.
  3. Review both diffs without their progress transcripts. Keep a surface only if all expected checks pass and the patch can be reviewed within your time cap. Halt the trial if setup requires unsupported privileges or credentials outside the approved environment.
  4. If you enable GitHub review, repeat on a separate PR with an intentionally mislabeled benign assertion. Track helpful and dismissed comments apart from coding completion.

Related articles