Using Codex mobile without downgrading your review
Published 2026-05-19 by Owner
Codex moved into the ChatGPT mobile app on May 14. This guide is about using it for what a phone is good at and not using it for the one thing a phone is bad at. For what shipped, see OpenAI brings Codex to the ChatGPT mobile app.
Who this is for
You already run Codex on a Mac and want to monitor or unblock runs while away from the desk. Skip this if you do not have a Mac running the Codex app — the mobile client is a remote control for that host, not a standalone agent.
Prerequisites
- The Codex app on the Mac, updated to the version that has the Codex mobile section
- The ChatGPT mobile app, updated, signed in to the same account
- Both devices on the same Wi-Fi network (or the phone on a VPN into it) so it can reach the host
Pair the phone
First confirm the host build:
codex --versionThen pair the phone. On the Mac, open the Codex app and select Codex mobile; it displays a QR code. In the ChatGPT mobile app, open Codex and scan it. The phone then loads live state from that Mac — the same projects, files, credentials, skills, and configuration. There is no separate cloud copy; the host is the source of truth. The Mac app’s Codex mobile panel shows current pairing state; treat that panel as authoritative for anything beyond the version, since the beta’s command surface is still moving.
Use it for monitoring and redirection
The phone is good at three things: watching the terminal output of a running task, stopping a run that has gone the wrong way, and answering a yes/no you already know cold so the agent stops waiting. All three reduce latency on a decision you have effectively already made. Turn on completion and needs-input notifications so the phone works as a pager for those moments.
The signal that it is time to stop a run from the phone is concrete: the terminal output drifts from what you asked for — the agent is editing files you did not expect, installing something you did not request, or repeating a failing step in a loop. You do not need to read the diff to catch this; you only need to notice the task is no longer doing what you described. When the output stops matching the request, stop the run from the phone and finish the diagnosis at a real screen rather than letting it continue until you are back at the desk.
Do not approve unread diffs from the phone
The one rule that matters: a diff you have not read does not get approved from a phone. Reviewing a change means reading it, predicting the next hunk, and holding the surrounding file in your head — none of which survives a thumb-scroll. Approving a low-risk command you understand (run the test suite, install a pinned dependency) is fine. Approving a code diff you have not actually read is the failure mode, and a small screen makes it the path of least resistance. Keep diff approval on a real screen. The risk framing is the same one in AI coding security modes: the gate only works if a human actually reads what passes through it.
Make the rule categorical, not situational. “I will only approve small diffs on mobile” erodes under exactly the conditions that make the phone tempting — you are busy and away from the desk. “No diff approval from the phone” has nothing to negotiate in the moment.
Notification hygiene and revoking a device
Keep only completion and needs-input notifications on; turn the chattier ones off so the pager stays meaningful. Treat a paired phone as a credential to the host — it can see the same files and secrets the Mac can. If you lose the phone or stop using the pairing, revoke it from the Codex app on the Mac and re-pair fresh rather than leaving an idle device authorized.
