Codex Mobile App: Monitor & Control Your AI Coding Agent from iPhone or Android (2026)

TL;DR: On May 14, 2026, OpenAI rolled Codex into the ChatGPT mobile app on iOS and Android, available in preview on every plan including Free. You scan a QR code from your Mac, and then you can review diffs, approve commands, switch models, and dispatch new tasks from your phone. The worker process still has to be macOS — Windows support is promised but undated.

Three months after Anthropic put Claude Code in your pocket, OpenAI shipped the same idea — except Codex Mobile only talks to a Mac. If you live in Windows or Linux, you are watching from the bleachers until "soon."

Codex itself is no longer a small experiment. OpenAI says more than four million developers use it weekly, and the missing piece for most of them was not raw capability — it was a way to glance at a running task while away from the desk. The mobile launch closes that gap.

What Codex Mobile actually does

Codex Mobile is a control surface for a Codex worker that runs somewhere else. The "somewhere else" is currently a Mac — your desktop, a Mac mini, or a remote Mac you have SSH'd into — and the phone is the window onto it. You do not run the model on the device. The device runs the review.

Concretely, OpenAI describes the experience this way: "From your phone, you can work across all of your threads, review outputs, approve commands, change models, or start something new." In practice that means four things:

1. Live observation. Screenshots, terminal output, diffs, test results, and approval prompts stream into the app in real time, with the worker's permissions and credentials staying on the host machine.
2. Approvals. When Codex hits a command that needs sign-off — running migrations, touching production config, deleting files — you get a card you can approve or reject from the lock screen.
3. Dispatch. You can open a new thread, attach context, choose between models (including the gpt-5.3-codex variant available through OpenAI's own product surface), and start work without going back to the Mac.
4. Thread continuity. Whatever Codex is doing on your desktop is reachable on your phone, and vice versa, including across multiple concurrent sessions.

This is not a tiny IDE in your pocket. It is the manager view of a coding agent — the part that used to require you to be back at your desk to unblock the agent on a permission prompt.

Setup: from QR code to first approved diff

The pairing flow is deliberately frictionless because it had to be — most developers will try this once on a coffee break, and a bad first run kills the habit.

On your Mac you launch Codex in whatever surface you already use (CLI, the desktop app, or the Chrome extension that shipped alongside it). Mobile linking lives in the same Codex settings panel. It prints a QR code. You open ChatGPT on the phone, point the camera at the screen, and the two halves of the same Codex session are now talking. There is no separate account, no extra API key, no provisioning step.

After that, every thread on the Mac is reachable on the phone within seconds. If you start a long-running refactor before walking to lunch, the agent's progress arrives as notifications — and any permission prompts surface as actionable cards.

It is worth noting how much of this design borrows from the Codex CLI workflow that has been the daily driver for a year now. The mental model is the same — agent runs, you supervise, you approve — but the supervisor seat has moved off the chair and onto the phone.

What you can actually do from your phone

Mobile is fine for a surprising amount of real work, and bad for a few things. The honest list, after using it for two weeks:

Genuinely useful on mobile:

• Approving migrations the agent has staged but won't run without sign-off.
• Reading a generated test failure and asking Codex to try a different fix.
• Watching a long build or test suite without keeping your laptop open.
• Spinning up a fresh thread from a screenshot or a bug report someone pasted in Slack.
• Switching the model mid-task — for example bumping a thread from a cheaper variant up to the strongest coding model when you realize the work is harder than expected.

Painful on mobile:

• Anything that requires reading more than ~80 lines of diff at a time. The pinch-to-zoom on a code block is workable, not pleasant.
• Heavy multi-file refactor planning. You will want a real screen for that.
• Anything that requires you to type more than a paragraph of clarification. Voice input helps, but only somewhat.

If you have spent any time choosing between Claude Code, Codex, Cursor, and DeepSeek-TUI, you already know the right framing here: each tool occupies a slightly different point in the who-watches-the-agent design space. Mobile pushes Codex further into the "asynchronous, supervisor-style" corner.

The honest limitations today

The preview ships with three real ceilings, and pretending otherwise would be silly.

It only talks to a Mac. Today, the worker has to be running on macOS. Windows is explicitly listed as "coming soon" by OpenAI, with no committed date. Linux is not mentioned at all in the launch material. If your daily driver is a Linux workstation, the only honest workaround right now is a Mac mini sitting somewhere on your network — which is not a small ask just for mobile parity.

Network dependence is harder than it looks on paper. The mobile experience leans on a persistent, low-latency connection to your worker host. Spotty café Wi-Fi turns "approve and continue" into "wait 9 seconds, approve, wait 7 seconds, see if it worked." Worth knowing before you plan a flight around it.

There is no offline mode. If your Mac is asleep or the host is down, the mobile app shows you the last known state and nothing else. Set your Mac to never sleep, or use Remote SSH with a server you control. (Remote SSH and Hooks both ship on every plan, per OpenAI's launch notes.)

These are not deal-breakers, but they are the difference between "I can use this on a real day" and "I demoed this once at a conference."

How it compares to Claude Code Remote Control

OpenAI is not first to this idea. Anthropic shipped Remote Control for Claude Code in February 2026, with broadly the same shape — phone observes, desktop works, approvals arrive as cards.

The differences that matter day to day:

Anthropic got there first with the cleaner cross-platform story. OpenAI replied with broader plan coverage (Free tier) and a narrower host requirement. Mobile is now table stakes for any serious coding agent. The choice has shifted from whether you adopt the pattern to which agent you trust enough to hand the lock-screen approvals to.

Where ofox.ai actually fits — and where it doesn't

Worth being precise about, because the question comes up every time a new ChatGPT feature drops.

Codex Mobile is not something you can wire to ofox.ai. It is the ChatGPT product, tied to your OpenAI account, billed through OpenAI's plans. There is no "use my API key instead" toggle. The mobile experience is a feature of the OpenAI consumer surface, not a model endpoint.

Where ofox.ai is genuinely useful around this: the same gpt-5.3-codex model that powers Codex inside ChatGPT is also accessible as an OpenAI-compatible API endpoint through ofox.ai's unified gateway. That matters if you want to build a different mobile experience — your own internal tool, a Telegram bot, a Slack workflow — where Codex-quality coding sits behind your own UI and you do not want to be locked to ChatGPT as the only client. The model is the same; what's different is who owns the wrapper.

If you are running Codex CLI through a custom provider configured for ofox.ai, the new mobile app does not affect your setup at all. Your CLI keeps doing what it does; the mobile feature is a parallel surface that just happens to share branding.

When mobile monitoring actually helps

Some honest use cases from the first two weeks of using this:

• Long migrations. Kick off a database refactor before a 90-minute meeting, approve the staged commands during the break, walk back to a green test suite.
• Cross-timezone handoff. Leave a Codex task running overnight, approve any blockers from a phone over breakfast.
• Code review while commuting. Reviewing the diffs Codex generated on an earlier task — not as good as on a 27-inch screen, but good enough to clear small ones.
• Production approvals. When a deploy script needs explicit human sign-off, getting that sign-off without having to be at your desk.

What it is not good for: writing the actual code. If you are using your phone to type instructions longer than two sentences, you are using the wrong tool. The mental model is "the agent works, you supervise" — and supervision is exactly the kind of work that a phone is genuinely fine for.

Your phone isn't replacing your IDE. It is replacing the chair you would otherwise have to sit in while the agent runs — and the developers who learn to trust that asynchronous loop will ship more than the ones still tethered to a desk.

Agentic coding has finally outgrown the assumption that the human has to be physically present while the agent works. That assumption was always a bit silly — nobody sits next to a CI runner watching it spin — and Codex Mobile is the first OpenAI product that admits it on the product surface. Whether you adopt it now or wait for Windows parity, the asynchronous loop it normalizes is where most coding work is heading.

Sources

• OpenAI: Work with Codex from anywhere — official launch post, May 14, 2026
• TechCrunch: OpenAI says Codex is coming to your phone
• 9to5Mac: OpenAI brings Codex to ChatGPT for iPhone and Android
• Engadget: OpenAI brings its Codex coding app to mobile

Originally published on ofox.ai/blog.