A coding task can need your attention while you are away from your computer. The agent may be waiting for a command approval, a decision about scope, or a review of the changes it made.
Codex in the ChatGPT mobile app lets you handle those parts from your phone. You choose a connected computer and project, send the task, follow its progress, and inspect the result. The code still runs on the selected computer.
In my demo, I use this setup with an image-generator project. I ask Codex to explain the code, suggest an improvement, write a plan, and implement it. Codex reports eleven passing offline tests, then I ask for a browser preview so I can check the interface too.
Below is the complete setup and review workflow, including the screenshot workaround for an app running at localhost. Start with a small change and a project you can verify.
The full video walkthrough is here.
https://youtu.be/CUIkNCYl2lI?embedable=true
Prepare the computer that will run the task
You’ll need:
- A ChatGPT account and workspace with Codex access.
- The latest ChatGPT desktop app on your computer.
- The latest ChatGPT mobile app, signed in to the same account and workspace.
- A computer that stays awake, online, and running the desktop app.
Get the apps from the official ChatGPT download page. Open Codex on your computer and make sure your project is available there before pairing.
OpenAI’s current remote setup guide lists Mac and Windows hosts, with iOS and Android clients depending on rollout. The phone steps here show my iPhone. Some app versions still call the feature Remote instead of Codex. A workspace admin may also need to enable Remote Control. Check the supported setup and requirements if your menus differ.
You start this setup in the desktop app. Installing only the Codex CLI or IDE extension won’t give you the pairing screen.
Pair the phone and keep the host available
On your computer:
- Open Settings, then Connections.
- Select Control this PC or Control this Mac.
- Enable Allow connections if it is off.
- Click Add. Your version may show Set up instead.
- Complete any verification and open the pairing QR code.
Scan your own code. The code in my recording is hidden.
On your phone, scan the code with the camera. Open the ChatGPT link, check the account and workspace, and approve the connection. Complete any passkey, SSO, or multi-factor authentication prompt it shows.
Each pairing connects a particular device to a particular computer. Repeat setup if you want another phone to control another host. You can manage paired devices in the computer’s Connections settings. Official pairing instructions
Enable Keep this PC awake, or the equivalent option, if available. Leave the desktop app open and your computer connected to power. A sleeping or offline computer can’t do the work you send it.
Now open Codex in the ChatGPT mobile sidebar. Your connected computer should appear. In the video, I also use Face ID or the device passcode to protect access on my phone.
Choose local or cloud execution deliberately
My mobile home screen has All, Cloud, and a filter for my computer, followed by saved chats.
Choose your connected computer for this walkthrough.
Select your computer to use its project files and installed tools. Cloud runs work in a separate cloud environment; it doesn’t automatically have the files sitting on your laptop. For that setup, see my Codex Cloud tutorial. Both options are available through the Codex mobile workflow, subject to account and workspace availability.
Start a new chat and select a project. You may also see Don’t work in a project, but selecting the folder you want to change makes this tutorial easier to follow.
Before sending anything, check the computer and project names. It’s easy to open a familiar chat and forget which machine it belongs to.
Set the permissions before the first prompt
Open the model controls to choose a model and reasoning effort. In my recording, I use High. You can begin with the available default; matching my exact model isn’t necessary.
Next, open the approval settings. My version shows:
- Ask for approval: asks before editing external files or using the internet.
- Approve for me: reviews actions automatically and asks about those detected as potentially unsafe.
- Full access: gives broader computer access and is marked as elevated risk.
The approval options shown in my demo. Host and workspace policies still apply.
I use Approve for me in the video. Choose the narrowest access that works for your task, and read a requested command before approving it. Automatic review doesn’t make every action harmless.
My first prompt is simple:
Tell me about this repository.
For your first attempt, you can make the request more specific:
Explain what this project does, how to run it, and how to test it.
Read its instructions first. Don't change any files yet.
Start by understanding the project before asking for changes.
You can follow progress and expand activity entries to inspect commands. In my recording, Codex looks for AGENTS.md, the file where a project can keep instructions for coding agents. It then summarizes the code and how to run it.
Turn a suggestion into a bounded task
My next question is:
How can we improve this repository?
Read the suggestions and choose one manageable change. Asking for every improvement at once makes the result much harder to review on a phone.
In the demo, I ask for a plan with an achievable goal. Codex proposes shared input checks and faster loading of saved image-generation history. The plan includes a goal, scope, checklist, and Done when conditions.
A small scope gives you something concrete to check afterward.
Here’s a reusable version of that request:
Pick one small improvement from your suggestions.
Create a plan with a goal, scope, checks, and a clear "Done when."
Explain what you will leave out. Don't implement it yet.
You can also enable Plan mode from the plus menu. Pursue goal is useful for an outcome that needs several rounds of work. A written checklist alone doesn’t mean you’ve started a persistent goal. My Codex goal guide explains that workflow.
Once the plan looks right, send:
Implement the plan. Run the relevant tests and tell me
what passed, what failed, and what still needs a manual check.
While Codex works, use Queue for a follow-up that can wait. Use Steer when you need to correct the current direction. For example, “Keep this change inside the settings page” belongs in the current work, not after a large rewrite. OpenAI explains these controls in its remote engineering guide.
Review the test report and the actual interface
The result in my recording has two parts: Codex reports that all 11 offline tests pass, and it says browser UI verification is still outstanding. Read both. A passing test report gives you evidence about the checks that ran; the unfinished browser review tells you what to do next.
Passing tests and checking the interface are separate parts of the review.
I ask Codex to start the project. It runs on the computer at localhost, using port 8765 in this example.
Opening that address on the phone doesn’t open the computer’s app. Localhost means the device opening the address. On your phone, it points back to your phone.
My workaround is to ask Codex for screenshots. Try:
Open the running app on the computer and take screenshots
at desktop and mobile widths. Show them here so I can review
the layout. Check the main interaction and report any errors.
In the recording, Codex uses a headless browser, a browser running without a visible window, to capture the app. The images appear in the chat, where I can open and zoom into them.
Check whether buttons fit, text is readable, and the layout works on a narrow screen. Screenshots show appearance; ask for interaction checks too. If Codex lacks browser tools on your host, it needs that setup before it can capture the app.
Make the commit a separate decision
Open Changes to inspect the diff: the lines added, removed, or edited. Use Files when you need more context. Ask Codex to explain anything you don’t understand before committing.
If you expected an existing repository, first check that you selected the correct folder. For a new project, open a terminal in that project’s root and run:
git init
git status
Before the first commit, check .gitignore so secrets such as .env files, dependencies, and generated output stay out of Git. Then ask Codex:
Review the diff and Git status. Commit only the changes for
this task with a short, clear message. Exclude secrets and
unrelated files. Don't push yet. Show me the commit hash.
A commit saves a checkpoint in your local Git history. Pushing sends commits to a remote repository such as GitHub. Deploying publishes the application. Treat them as separate steps, so you know exactly what you’ve approved.
Keep follow-ups and reviews easy to find
Type / in the composer to see the commands your version supports. The most useful starting points are /plan for planning, /review for checking changes, /side for a side question, and /status for session details. Availability depends on your app, account, and host. Mobile command reference
Attach a screenshot when describing a visual problem. “This button overlaps the title” is much easier to act on when Codex can see it.
Rename chats around the task, pin active work, and archive finished chats from the three-dot menu. A name like “Fix mobile menu” is easier to find later than “New chat.”
If the connection doesn’t work
If your computer is missing, check that the desktop app is running, connections are enabled, and both devices use the same account and workspace. Update both apps and pair again if needed.
If a session disconnects, check the computer’s power and network first. Keeping the phone awake won’t help if the host has gone to sleep.
If you can’t find Codex or Remote, check feature availability and your workspace’s Remote Control settings. Signing out can also turn Remote Control off; enable it again after signing in. OpenAI keeps the connection troubleshooting steps in its setup guide.
The part I would repeat is the review sequence: give Codex a specific task, check its plan, read what it tested, and inspect the result before committing. On a phone, keep the change small enough that you can make those decisions carefully. If a diff needs a larger screen, leave that review for your computer.
Cheers, proflead 😉