Codex on my phone: steering agent tasks while the Mac works

I steer from the phone, Codex works on the Mac, and I check the result.

1. Start or steer from the phoneOpen the same Codex task, give direction, answer questions, review progress and approve the limited actions it asks about.
2. Execute locally on the MacThe desktop app keeps access to the repository, terminal, browsers and the current working state.
3. Check the resultRead the diff, screenshots, tests, deploy result and live check before I call the task done.

Remote Codex continues the task on the Mac

Remote Codex lets me continue the same agent task from my phone while the Mac does the work. The Mac holds the source, tools, credentials and browser state, so the task runs there.

Unlike remote desktop, the phone does not mirror the Mac’s screen. Codex shows me the task’s progress, its questions and the actions waiting for my approval.

The setup I use

1. On the Mac, I open the project in the Codex desktop app and sign in to the same account and workspace I use on my phone.
2. I enable remote access for that desktop session and leave the app running.
3. I keep the Mac awake, connected to the internet and able to reach the tools the task needs.
4. On the phone, I open Codex, choose the connected desktop environment and continue the task.

The host has to stay awake, online and running the Codex desktop app. If any of the three fails, the phone has nothing to connect to. OpenAI’s remote connections guide was the setup reference when I published this note on 25 July 2026.

Keeping the Mac available with Amphetamine

I use the macOS utility Amphetamine to stop idle sleep during a remote work window. I set a fixed duration or a clear trigger, keep the charger connected for long sessions and end the session when the work is finished.

So I count on Amphetamine only to prevent idle sleep. A closed, unplugged MacBook may still go to sleep and drop the connection, so I test the physical setup I plan to use before I depend on it.

How the work moves between Codex and me

Remote tasks go best when I split them into checkpoints. I give the outcome and say what Codex may change. Codex checks the live site, makes the change within those limits and shows me what it did. Any decision that would change the scope, publish something, spend money or affect another person comes back to me first.

1. BriefOne outcome
2. BoundaryWhat may change
3. CheckpointWhat needs approval
4. ProofWhat closes the task

This page went through that loop. I asked for the story, and Codex and I checked the current live design instead of trusting an old template. We saved its typography and colour system as notes for later sessions, built the page, deployed it and tested the live result on desktop and mobile.

Prompts I send from the phone

Inspect first. Tell me which production source is current.

Continue the same task. Do not widen the scope.

Show me the exact copy before publishing.

Deploy to production, then test the live URL on mobile
and desktop. Submit the changed URL to IndexNow only
after the live checks pass.

Short prompts work because the task already carries its files, decisions and history. The thread keeps enough of that context for me to switch devices without starting again from zero, though it does not keep everything.

Guardrails that matter more in remote sessions

• Use the same account and workspace. That is what links the phone to the desktop environment.
• Keep the host physically secure. An awake, unattended computer can reach valuable systems.
• Give each wake session an end. Do not leave the Mac awake indefinitely by default.
• Keep secrets out of prompts and anything I publish. Working remotely does not change how credentials are handled.
• Close the task on the live check. An agent’s “finished” message is not proof that the live result works.

Suhail Mohebi · published 25 July 2026 · the cover is a generated illustration