Codex remote control: driving your dev machine from your phone
Codex Remote pairs the ChatGPT mobile app to a host computer by QR code: start work, approve commands, review diffs from anywhere. Setup, what runs where, and what happens when the host sleeps.
There’s no separate Codex mobile app. Codex lives in the Remote tab of the ChatGPT app for iOS and Android. You pair your phone to a host computer with a QR code (Settings → Connections → “Control this Mac or PC” on the host), and from then on you can start work, send follow-ups, approve commands, and review diffs from anywhere. Generally available on every ChatGPT plan, including Free, since June 2026. The catch is structural: the host computer does all the work, so if it sleeps, goes offline, or closes the app, remote access stops until it’s back.
What Remote actually is
Codex Remote splits the work in two. Your phone sends prompts, approvals, and follow-up messages; the connected host provides everything Codex actually uses: the files, the shell, MCP servers, skills, browser automation, and credentials. From the phone you can start new chats in projects on the host, steer running work, approve commands and other actions, review “outputs, diffs, test results, terminal output, and screenshots,” get notified when something needs you, and switch between connected hosts. Since late July 2026 you can even hold a voice conversation with a running session on iOS.
Transport is worth understanding because it’s better than the DIY guides assume: OpenAI runs a secure relay layer between your paired devices, so the host is never exposed to the public internet and you don’t need a VPN or mesh network for the phone-to-host path. (People do pair this setup with Tailscale for their own SSH access. That’s a sensible habit, but it’s community practice, not part of the feature.) Pairing is per-device-per-host: each phone scans a QR code for each host it should control.
Setting it up
On the host (macOS or Windows), the current ChatGPT desktop app, which absorbed the standalone Codex app after the July 2026 merger, carries the switch: Settings → Connections → Control this Mac or PC, then Set up. On your phone, the latest ChatGPT app, signed into the same account and workspace, scans the QR code. Requirements beyond that: the host stays awake, online, and signed in; your plan has Codex access (all plans do); and on Business/Enterprise workspaces an admin may need to enable Remote Control access. A desktop can also control other machines (“Control other devices”) where the feature has rolled out, and SSH hosts from your ~/.ssh/config can be added: the app starts a remote Codex process over SSH.
The hub problem, stated by the docs themselves
OpenAI’s documentation is refreshingly direct about the weak point: “If that computer sleeps, loses network access, or closes the app, remote access stops until it’s available again.” The suggested mitigations escalate exactly the way you’d guess: keep it plugged in, enable the keep-awake setting, leave a Mac’s lid open (closed-lid needs an external display), and finally: use “a dedicated always-on computer.” That last line is the whole hub-and-spoke pattern conceding its dependency. A laptop that follows you around is a bad server; people who lean on Remote seriously end up pointing it at a Mac mini on a shelf or an always-on box that exists to run agents, a machine whose job is to never be asleep.
Codex app-server, briefly
Under the hood, rich Codex clients talk to codex app-server, an open-source JSON-RPC interface (stdio, WebSocket, or unix socket) that exposes authentication, conversation history, approvals, and streamed agent events. It’s what powers the VS Code extension, and you can run the CLI against a remote app-server over WebSocket. Two things the docs say before you build on it: OpenAI marks the command and its WebSocket transport experimental (“not supported for production workloads,” and non-loopback listeners currently accept unauthenticated connections during rollout), and the docs explicitly tell you to reach remote machines through a VPN rather than exposing an app-server to the internet. The pattern making the rounds (host app-server in the cloud, let users sign in with their own ChatGPT subscription) extrapolates from real pieces, but it isn’t a documented, supported shape today.
Phone-first coding, compared
If you’re choosing a phone-driven setup: Codex Remote keeps execution on your own hardware with OpenAI relaying the control channel; Claude Code’s phone story offers both that shape (Remote Control) and true cloud sessions where Anthropic hosts the VM. Either way the pattern converges on the same requirement: a computer that’s reliably on when you’re not next to it. The phone was never the hard part.