Codex remote control: driving your dev machine from your phone
Codex Remote pairs the ChatGPT phone app with a Mac or Windows host by QR code. Setup, plans, codex remote-control, and what happens when the host sleeps.
There’s no separate Codex phone app. Remote control lives in the ChatGPT app for iOS and Android: on iOS it is now called Codex, and an app that hasn’t updated still shows Remote. You pair the phone with a Mac or Windows PC running the ChatGPT desktop app by scanning a QR code (on the computer: Settings → Connections → “Control this Mac or PC” → Set up), and from then on you can start tasks, steer them, approve commands and review diffs from anywhere. It has been generally available since June 25, 2026, and OpenAI’s plan table lists it on Plus, Pro, Business, Enterprise and Edu. The catch is structural: the computer does all the work, so if it sleeps, goes offline or closes the app, remote access stops until it’s back. The terminal command codex remote-control is a separate, experimental piece; the documented setup starts in the desktop app.
What Remote actually is
Codex Remote splits the work in two. Your phone sends prompts, approvals and follow-up messages; the connected computer, the host, provides everything Codex actually uses: its projects, files, shell, credentials, MCP servers, skills, plugins, browser setup and Computer Use. From the phone you can start new chats in the host’s projects or continue existing ones, steer running work, approve commands and other actions, review “outputs, diffs, test results, terminal output, and screenshots,” get notified when a task finishes or needs you, and switch between connected hosts. The host’s sandbox settings and approval rules still apply to everything started from the phone. ChatGPT Voice works through it on iOS on Plus, Pro, Business, Edu and Enterprise; talking inside a task that began as typed messages is still rolling out.
OpenAI runs a secure relay layer between your paired devices, so the host is never exposed directly to the public internet and the phone-to-host path needs no VPN or open port. Pairing is per device and per host: each phone scans a QR code for each computer it should control. Connections used since June 8, 2026 stay paired; an older one needs both apps updated and a fresh pairing.
Setting it up
The host is a Mac or Windows PC running the current ChatGPT desktop app, which absorbed the standalone Codex app on July 9, 2026. Open Settings → Connections → Control this Mac or PC, select Set up (or Add), approve remote access and complete any verification it asks for. Scan the QR code with your phone, confirm the same ChatGPT account and workspace, and finish any multi-factor, SSO or passkey step; the host then appears under Codex in the app. OpenAI’s setup page is plain that this starts in the desktop app: you can’t set it up from the Codex CLI or the IDE extension. The Linux desktop app, in preview since August, is not among the hosts it lists.
Beyond that, the host stays awake, online and signed in, and your plan has to include the feature. OpenAI’s plan table lists mobile remote control on Plus, Pro, Business, Enterprise and Edu, and not for Codex signed in with an API key; the Free and Go plan cards mention only the desktop app. In a workspace, an admin may need to enable Remote Control access first. Signing out of ChatGPT on the host turns Remote Control off without deleting your pairings, so turn it back on after you sign in again.
A desktop host can also work in a remote environment. Add the server to ~/.ssh/config, check that ssh reaches it from the host, install and sign in to Codex on the server, then enable it under Settings → Connections and choose a project folder. The app starts Codex’s app server on the server over SSH through your login shell, so codex must be on that shell’s PATH. Your phone still connects to the desktop host, which runs the work on the server. For a server outside your network, OpenAI’s advice is a VPN or mesh network rather than exposing the app server to the internet; Tailscale for SSH is one such setup. Where it has rolled out, a desktop host can also control another Mac or PC through Control other devices.
The codex remote-control command
The CLI carries a command with the feature’s name, and it is a different piece. In Codex CLI 0.160.0, released October 1, 2026, codex remote-control runs remote control in the foreground, start and stop run it as a background app-server daemon, and pair prints a short-lived manual pairing code:
codex remote-control start
codex remote-control pairCopiedOpenAI’s command reference labels it Experimental and says managed remote-control clients and SSH remote workflows use it. The iOS app has accepted manual pairing codes since June, but OpenAI’s Remote setup still starts every pairing in the desktop app and lists only Mac and Windows hosts, so a phone paired straight to a headless Linux server running this command is outside what the docs cover. A July 2026 report against CLI 0.142.5 described exactly that kind of server becoming impossible to add once pairing moved to QR codes, because the CLI had no pairing command then; current releases have one.
The hub problem, stated by the docs themselves
OpenAI’s documentation is 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 the way you’d guess: keep it plugged in, turn on keep-awake in the host’s connection settings, leave a Mac laptop’s lid open with power connected (closed, it also needs an external display, and choosing Sleep still stops it), keep a Windows host unlocked for Computer Use tasks, and finally add “an always-on computer or SSH host when you need continuous access.” That last line is the hub-and-spoke pattern conceding its dependency. A laptop that follows you around is a bad host; for Codex Remote the role wants a Mac or PC that never sleeps, such as a Mac mini on a shelf.
Moving the project onto a server does not take the desktop host out of the path. OpenAI’s docs describe the phone connecting to the desktop host, which works in the SSH environment, and they don’t describe reaching the server with that host asleep. On the docs as written, the machine that has to stay awake is the Mac or PC, wherever the code lives.
OpenAI’s other answer leaves your computer out of it. Codex Cloud tasks run in an isolated workspace OpenAI hosts, started from an environment you prepare from GitHub repositories and publish on the web or in the desktop app; since DevDay on September 29, 2026 those environments are reusable, and on the phone the same Codex entry offers them. They “keep working while your computer is asleep.” What they don’t have is your computer: its local files, running processes, browser sign-ins and VPN access aren’t carried over. The plan table lists Codex Cloud on Plus and up. DevDay’s “work across devices” is a different feature, for ChatGPT Work tasks, and the docs say it does not change Codex Remote.
Codex app-server, briefly
Under the hood, rich Codex clients talk to codex app-server, an open-source JSON-RPC interface that exposes authentication, conversation history, approvals and streamed agent events. It powers the VS Code extension, the desktop app starts it on an SSH host, and since the separate Codex MCP server was removed on September 5, 2026, it is the integration route OpenAI points to. It talks over stdio by default, or a WebSocket or Unix socket, and the CLI can attach to one on another machine with codex --remote. Two things the app-server docs say before you build on it: the command and its WebSocket transport are experimental and “aren’t supported for production workloads,” and a WebSocket listener on anything but loopback still accepts unauthenticated connections by default during rollout. They give token flags to set first (--ws-auth capability-token or signed-bearer-token), want TLS for anything beyond localhost or an SSH tunnel, and for a machine outside your network recommend a VPN or mesh network over exposing the app server to the internet.
The pattern that made the rounds this summer, an app-server in the cloud that users sign into with their own ChatGPT subscription, now has an answer in the docs. The app-server page says its authentication “has never been permitted for commercial or hosted services” and points those at Sign in with ChatGPT. Its plan-usage option, announced at DevDay, lets an open-source or locally hosted app run Codex app-server on a Plus or Pro user’s own plan, a self-hosted VM included, as a preview open to open-source partners and selected private clients; a paid or remotely hosted app goes through an interest form.
Phone-first coding, compared
If you’re choosing a phone-driven setup, both vendors now offer the same two shapes. Codex Remote keeps execution on your own computer with OpenAI relaying the control channel, and Codex Cloud runs the task on OpenAI’s machines; Claude Code’s phone story pairs Remote Control with cloud sessions on Anthropic’s. The difference that matters for an always-on server is where the host side starts. Claude Code’s Remote Control starts from the terminal, where claude remote-control keeps running as a server, so a headless Linux box can be the host; Codex’s documented host is the desktop app on a Mac or PC, with a Linux server reached through it over SSH. For work that needs your own files, tools and sign-ins, both still need a computer that stays on while you’re away from it.