Article

Codex Mobile Bridge: Your AI Coding Agents, Now on Your Phone

Codex Mobile Bridge: Your AI Coding Agents, Now on Your Phone

Your coding agent is halfway through a refactor. It needs you to approve a file edit, answer a clarifying question, or confirm that yes, it may run the test suite. You are not at your desk — you are on a train, in a queue, or on the couch with your phone — and the agent is waiting.

Codex Mobile Bridge is an open-source project that solves exactly that gap. It runs a small gateway on your computer, and that gateway turns the AI coding tools already installed there — Codex, Claude Desktop, DeepSeek Harness, ZCode, and, since the latest release, Google's Antigravity — into a workspace you can drive from a phone browser, an Android app, or an iPhone. The code keeps running on your machine; the phone is only the control surface. The project, built by a Chinese developer and actively discussed on LINUX DO, has been iterating fast: version 2.2.1 landed with a fifth client and a reworked connection system.

Five agents, one phone screen

The Bridge is not itself an AI model or a coding agent. It is a remote control layer that sits in front of the clients you already use, showing you the same sessions, projects, and chats that exist on your desktop — because it talks to the native clients rather than re-implementing them.

As of v2.2.1, the supported clients are:

ClientWhat carries over to your phone
CodexLocal and SSH sessions, models, Skills, approvals, plan and goal modes, side chats
Claude DesktopCode and Cowork sessions, grouped; supported approval responses
DeepSeek Harness (DSH)Chats via the Harness plugin, model and quota views, a native mobile page
ZCodeNative chats, attachments, approvals, model and Skill selection, API quota
AntigravityApp, IDE, and CLI history, native questions, child-task management (new in v2.2.1)

That breadth is the project's real selling point. Most remote-control tools are built for one agent; this one treats your desktop as a small fleet and gives you a single switchboard. Sessions stay separated per client, so a Codex chat and a Claude Cowork session never bleed into each other.

How the gateway actually works

The architecture is deliberately boring, which is a compliment. You install a gateway on the computer that runs your agents — macOS, Windows, or experimental Ubuntu on x64 and ARM64 — and it serves a web workbench. Your phone reaches it over the local network, or through a Cloudflare temporary or fixed HTTPS tunnel, Tailscale, an SSH hop to your own server, or a NAS.

Setup follows four steps, per the project's own guide: install and log in to the clients you need on the computer, open the Bridge gateway and scan for installed clients, scan a QR code from your phone to log in, then pick a chat and continue. The QR code is a short-lived, one-time login credential, and the gateway has its own independent login — you do not hand your OpenAI or Anthropic account to the phone. Model calls still use whatever authentication the desktop client already has.

Two design choices matter. First, the gateway bundles its own runtime, so first local-network setup needs no Python, no Node.js, and no agent deployment. Second, everything still executes on the host: if the computer sleeps, disconnects, or powers off, the remote session stops with it.

What you can actually do from your phone

A chat window on a phone is the easy part. The interesting question is how much of a real workflow survives the shrink, and the answer is: more than you would expect.

Files follow the project. You can browse the project directory, pull a file or folder from the computer into a message without downloading it to the phone first, and preview Markdown, source code, and images. Chat downloads have an adjustable cap, so a stray 200 MB log cannot eat your data plan by accident.

The terminal is a real terminal. This is not a "run one command" box. The Bridge keeps a persistent shell session — cd survives, environment variables survive, interactive state survives — using your actual login shell (zsh on macOS, bash on Linux, COMSPEC on Windows). On a phone, newlines in the input box are for composing; you tap send to execute, which avoids running a half-typed command by accident.

Git review works, with a guardrail. You can inspect diffs, stage files, write a commit message, create or switch branches, and merge locally. What you cannot do is push to a remote, rewrite history, or force-discard changes. That is a deliberate limitation, and a sensible one: a phone is a bad place to discover you just force-pushed over your team's main branch.

Approvals and questions are the core loop. The whole point of the tool is the moment an agent pauses and asks. From the phone you can respond to command, file, and question cards, approve a computer-use scope, skip individual async questions in Codex without losing your other answers, and watch subagent progress in a task tree (read-only for now).

Version 2.2.1 adds one more flourish: interactive charts embedded in Codex replies, rendered with sliders and buttons you can actually use, with their state restored when you reopen the chat.

Notifications tie it together. Android can surface per-event background notifications (off by default), iOS leans on foreground sync plus Live Activities, and both can route through Bark, ntfy, or PushPlus. The project is honest about the limits: iOS background delivery is not guaranteed, and Android notifications depend on system permissions and battery policies.

The rough edges, stated plainly

No tool in this category is frictionless, and the project's own documentation is unusually candid about where it bites.

iOS is the awkward sibling. There is no App Store or TestFlight build. The IPA must be self-signed with your own Apple account, either directly or by building the Xcode project. That is a real setup cost, and it recurs on updates. Android is simpler — an APK you install directly, with in-app updates — but v2.2.1 introduces a wrinkle: the new Bridge app installs alongside the old v2.2.0 app rather than replacing it, so you must re-add your computers and manually preserve any drafts before uninstalling the old one.

The host is a single point of failure. Your computer has to stay awake, online, and running both the client and the gateway. SSH sessions need their hosts reachable too. If you want to code from your phone while your laptop sits closed in a bag, this is not the tool — you need a machine that stays on, which is why many users run it on a desktop or a home server.

Some things still need the desktop. Complex MCP forms, identity verification challenges, and certain special confirmations still require the real client. Claude's plain Chat is not supported yet (Code and Cowork are), and WSL account management remains experimental — though v2.2.1 improves WSL status reporting and adds a "start and connect WSL" entry.

The Mac build is unsigned. The macOS installer is not notarized, so the first launch needs a manual allow in System Settings, and the Windows build is not code-signed either. The releases ship SHA-256 checksum files, and verifying them before installing is the sensible move.

Where it fits in a crowded field

The idea of controlling a coding agent from a phone is no longer exotic. OpenAI and Anthropic both ship official remote-control features for their own agents now, and several independent projects cover the same ground for Codex and Claude specifically. Codex Mobile Bridge's bet is different: instead of going deep on one agent, it goes wide, and instead of routing you through a vendor's cloud, it keeps the whole loop on hardware you own.

That self-hosted stance is the reason it resonates with the LINUX DO crowd. The forum thread has been running since late September 2026, gathering 127 posts and steady updates from the maintainer, and the discussion reads less like a product launch and more like a public changelog. It is the same pattern that made self-hosted remote desktop a fixture of Chinese dev culture: if the tool touches your code and your credentials, you would rather run it yourself. Readers who like that philosophy will recognize the kinship with the self-hosted remote desktop stack Chinese developers already use for reaching office PCs — and with RustDesk, the open-source remote desktop that plays the same "your server, your rules" game for full desktops.

Should you try it?

If you already run Codex or Claude Desktop on a machine that stays powered on, and you regularly find yourself away from the keyboard while an agent waits on you, the answer is yes — with the setup cost honestly counted. Start on your local network, where setup needs nothing beyond the gateway and a QR scan, and only add a remote route (Tailscale is the least fiddly; a Cloudflare tunnel needs no domain) once the local flow feels solid. Keep the gateway and the mobile app on the same version; the project warns that mismatched versions are the most common source of weird behavior.

If you use exactly one agent, only occasionally, and your machine is a laptop that sleeps when you close it, the official remote-control features from the agent vendors will probably serve you better with less maintenance. The Bridge earns its keep when you juggle several agents, want one phone screen for all of them, and prefer a gateway you can inspect, update, and kill yourself to a cloud service standing in the middle.

Keep reading
One HTML File, One Full 3D Shooter: What an AI-Built Game in 432 KB Really Means
Read next →