Codex used to require a desktop. That changed.

OpenAI quietly expanded Codex access to the ChatGPT mobile app — iOS and Android — in May 2026. You can now hand off a coding task from your phone, lock the screen, and come back to a pull request. The agent does the work in the background while you do something else.

That is useful. It also takes about three minutes to break if you skip the setup. Here is what you actually need to do.

What Codex is and what it is not

Codex is not autocomplete. It is not the coding model that ships inside ChatGPT's chat interface.

Codex is a cloud-based software engineering agent powered by codex-1, a model OpenAI built specifically for software tasks. When you send Codex a task, it spins up an isolated cloud sandbox, seeds it with your GitHub repository, and runs: writing code, editing files, executing tests, and opening pull requests. You are not in the loop until it is done.

The distinction matters because AI agents and chatbots operate differently. A chatbot answers. An agent acts. Codex takes actions inside your codebase — that is the whole point, and also the reason you should not point it at a repository you have not thought through.

What you need before you start

Three requirements, non-negotiable:

  1. A ChatGPT subscription at Pro, Team, Enterprise, or Edu tier. Free and Plus accounts do not have Codex access as of May 2026.
  2. A GitHub account with a repository you want Codex to work in. Codex requires a connected repo. It does not work against local code or arbitrary file uploads.
  3. The ChatGPT iOS or Android app, updated. Codex in mobile requires a recent app version. If you do not see it, update first.

That is it. No API key setup. No separate Codex dashboard. It runs through ChatGPT.

Connecting your GitHub repository

You do this once. After that, any Codex task you assign can pull from that repo.

  1. Open the ChatGPT app on your phone.
  2. Tap the model selector at the top — switch to Codex if it is not already selected.
  3. You will be prompted to connect a GitHub account the first time. Tap Connect GitHub and authorize OpenAI's GitHub app through the OAuth flow.
  4. Select the repository you want Codex to access. You can add more repositories later.
  5. Codex will confirm the connection. You are ready.

OpenAI's GitHub app requests read and write permissions on the repositories you select. Do not authorize it against repositories with production secrets, credentials in plaintext, or code you are not comfortable having an AI agent modify autonomously. This is not hypothetical caution — Codex can and will open pull requests.

Assigning a task from your phone

This is the part that actually changes your workflow.

Open ChatGPT, select Codex, and describe the task the way you would describe it to a junior engineer. Be specific:

  • "Add input validation to the email field in components/ContactForm.tsx. It should reject inputs without an @ symbol and show an inline error message."
  • "Write unit tests for the calculateDiscount function in lib/pricing.ts. Cover edge cases: zero price, negative discount, discount above 100%."
  • "Find all console.log statements in the /api directory and remove them. Do not touch console.error."

Vague tasks produce vague results. Codex does not ask clarifying questions mid-run the way a chat model does. What you type is the spec.

After you submit, Codex queues the task and starts working in a sandboxed environment based on your connected repo. You will get a notification when it is done.

What happens while it runs

You can leave. That is the point.

Codex runs asynchronously. It does not need you to stay in the app. The cloud sandbox isolates its work from your actual repository until it is ready to open a pull request. You will not see commits appearing in your repo mid-run — you will see a pull request appear when Codex finishes.

Review the PR before you merge it. Codex is good, not infallible. It can introduce logic that compiles cleanly but misunderstands the intent of a function. The PR is the handoff point. Use it.

What Codex actually handles well

Based on how codex-1 is designed and what OpenAI has documented:

  • Bug fixes with clear reproduction steps — give it the error message and the file, and it usually lands correctly.
  • Adding tests for existing functions — especially when the function behavior is already well-defined.
  • Refactoring isolated modules — works well when the scope is tight. Struggles when the refactor has implicit dependencies across the codebase.
  • Removing or migrating deprecated patterns — find and replace at a code level, with awareness of context.

Where it struggles: tasks that require understanding product intent, tasks that span many files with complex interdependencies, and anything where the right answer requires organizational knowledge that is not in the code.

What you should actually do

Connect it to a low-stakes repository first. A side project, a personal tool, a test repo. Run a few tasks. See how it handles your codebase before you route real work through it.

Review every pull request it opens before merging. This should be table stakes for any AI coding agent, but mobile use makes it easy to tap "merge" on a phone without reading the diff. Do not do that.

Keep the repo connection scoped. You do not need to authorize every repository you own. Authorize the ones you are actively working in.

Codex on mobile is genuinely useful for the window between desk sessions — delegating a fix before a meeting, queuing a test suite while you commute, handing off a refactor before you go to sleep. That is real leverage.

It is not a replacement for understanding what is getting merged into your codebase.