Remote Control or Cloud: Coding From Your Phone (2026)

There are two honest ways to do remote-control coding from your phone in 2026, and they are not the same product even though the apps look alike. The first is a relay: the AI agent runs on a computer you own and keep awake, and the phone drives it over an encrypted link — this is Happy Coder, Claude Code Remote Control, AgentsRoom, RemoteCode, VibeTunnel, and the Goose iOS app. The second is a cloud container: a hosted Linux box runs the agent for you, so nothing on your desk has to stay on — this is what we built with Cosyra. The relay route is usually free or a few dollars a month and keeps your code on your own machine; the cloud route costs money and needs a network but asks you to keep nothing awake. Sign up for Cosyra gets you 1 hour free, no credit card.

This is the pillar page for the "drive an agent from my phone" topic, and it sits under the broader AI coding agents on mobile pillar. If you want the agent itself walked through, see Claude Code on phone; if you care about the terminal layer underneath, see the mobile coding terminal guide. For the head-to-head pages on the individual relay tools, we have Cosyra vs Happy Coder, Cosyra vs Goose, Cosyra vs VibeTunnel (the tunnel that turns your own Mac into the remote box), Cosyra vs Moshi (the SSH and Mosh terminal app that drives an agent over your own connection), and Cosyra vs Catnip (the open-source iOS app that points Claude Code at a Codespace); for the cloud side, Cosyra vs Claude Code on the web weighs a container against Anthropic's own managed VMs. This page stays one level up: it explains the two architectures so you can tell which kind of tool you actually want before you install anything.

What does remote-control coding from your phone actually mean?

The phone is not where the work happens. On every option here, the AI model and the agent process run on a real computer with a real filesystem, and the phone is a screen and a keyboard pointed at it. What differs is whose computer. That single fact splits the whole category in two, and getting it straight saves you from installing the wrong kind of app.

In a relay setup, the computer is yours: a laptop on your desk, a desktop in the corner, or a home server. You install a small client on that machine, it launches a normal local agent session, and a phone app connects to it through an encrypted relay so you can read output and send input from anywhere. The compute, the files, and your API tokens stay on your hardware. In a cloud container setup, the computer is one a provider runs for you: you open an app, a Linux box spins up in the cloud, the agents are already installed, and the phone is a full client to that box. Nothing of yours has to be awake.

The honest opinion we will plant a flag on, and one the relay crowd will argue with: for most people, "keep a laptop awake at home so my phone can reach it" is a worse deal than it sounds. The day it matters — you are on a train with twenty minutes and an idea — is exactly the day the machine has gone to sleep, installed an update, or dropped off the Wi-Fi. We built a cloud box precisely because the always-on home machine kept failing us at the moment we wanted it. Plenty of careful developers disagree, and for good reasons we get into below.

Architecture 1: relay to a machine you own

A relay app gives you a phone client and an encrypted pipe to an agent running on your own computer. The selling points are real: it is usually free or a few dollars a month, your source code never leaves machines you control, and you reuse the Claude or Codex subscription you already pay for. The shared cost is also real: the host machine has to stay powered, awake, and reachable, or there is nothing for the phone to talk to. Five tools own this space, and they differ more than they look.

Happy Coder: the free, open-source relay

Happy Coder is open source (MIT) with about 21,800 GitHub stars as of 2026-06-12. You run its CLI wrapper on your own machine in place of claude or codex, and the iOS, Android, or web app drives that session. The relay is end-to-end encrypted and self-hostable, so you can keep code on devices you fully control, and it has realtime voice control, which we do not. It is genuinely free software. The constraint is the architecture: it provides zero compute, so a sleeping laptop strands it. Our full Happy Coder comparison goes deeper on where each side wins.

Claude Code Remote Control: first-party, Claude only

Remote Control is Anthropic's own feature (a research preview shipped 2026-02-25) that lets the Claude app or claude.ai/code drive a Claude Code session on your machine. It is included with a paid Claude plan — Pro, Max, Team, or Enterprise — so it is not free-tier and it does not work with an API key. It drives Claude Code only, not Codex CLI, Gemini CLI, or OpenCode, and the local process must keep running: the docs note an extended network outage of roughly ten minutes ends the session. If you live entirely inside Claude Code and already pay for it, this is the leanest relay there is.

AgentsRoom: the multi-agent orchestrator

AgentsRoom is the heaviest of the five: a desktop app (macOS, Linux, Windows) that spawns real CLI processes, each isolated on its own git worktree, with native iOS and Android companions over a TweetNaCl end-to-end-encrypted relay. It runs seven providers with mid-conversation switching, role-based agents, a kanban backlog, and per-agent diff review. It is free for up to 3 projects and $9.99 a month for unlimited, BYOK with no token markup, as of 2026-06-27. If you want to orchestrate several agents in parallel and review their diffs from your phone, nothing else here matches it — but the desktop command center has to stay awake. Our AgentsRoom on your phone guide walks through how its mobile apps pair to that desktop and what a no-desktop cloud-container setup gives up in exchange, and our Cosyra vs AgentsRoom comparison puts the orchestrator-on-your-own-desktop model head to head with a hosted box, feature by feature.

RemoteCode: the lean Mac-to-iPhone relay

RemoteCode is the minimalist option: an iOS/iPadOS app that pairs with a macOS desktop client to drive a local Claude Code, Codex, or OpenCode session. It is free on your own Wi-Fi or VPN and $4.99 a month for an internet relay, the cheapest paid tier in the group, and code never leaves your Mac. Two sharp limits, both verified 2026-06-27: the desktop client is Apple-Silicon macOS 15+ only, and there is no Android client at all, so an Android user cannot use it. Our RemoteCode comparison covers the relay-vs-cloud trade-off in full.

Goose: open source, with an iOS remote

Goose is an open-source (Apache-2.0) agent, originally built at Block and now an Agentic AI Foundation project, with about 47,300 stars as of 2026-06-07. The agent runs on your own machine; the free "Goose AI" iOS app tunnels into it over Cloudflare. The thing to get right is that the iOS app is a remote, not an on-phone agent — the model never runs on the phone — and there is no Android client shipped yet, only a work-in-progress port. If you already self-host Goose, the iOS app is a clean way to check on it. Our Goose comparison covers the self-hosting trade-off in full.

Onepilot: the iPhone SSH client for your own box

Onepilot is the same relay idea aimed squarely at iPhone: an SSH client that deploys Claude Code or Codex onto a Linux box you already run, with Telegram and Discord triggers on top. It is iOS-only (iOS 18.6+), free to start with a paid Pro tier, and like every tool in this section it needs your own server kept awake. Our Onepilot comparison covers where that BYO-server model beats a cloud container and where it does not.

Architecture 2: a cloud container does the work

The other architecture moves the computer off your devices entirely. Instead of relaying to a machine you keep awake, you open an app and a Linux box runs in the cloud. We run Cosyra this way: a per-user Ubuntu 24.04 x86_64 container on Azure AKS, with Claude Code, Codex CLI, OpenCode, and Gemini CLI pre-installed, reached from native iOS and Android apps and the web. It keeps 30 GB of persistent storage per user, and containers hibernate after about ten minutes idle and resume where they left off, so closing the app does not cost you the session. It is BYOK: you bring your own Anthropic, OpenAI, or Google key.

The reason we went this way is the always-on machine problem. With a cloud container there is no laptop to leave running, no home server to patch, no port forward that dies when your ISP rotates the IP. You can start a session from a phone whose owner has no other computer at all. The same container is reachable from an iPhone, an Android phone, and the web, with files intact, so you are not tethered to one driving machine the way a relay tethers you to its host.

We will name the trade-offs plainly, because they are the relay camp's best arguments. A cloud container is not offline: no network means no terminal, while a local relay can at least keep a session you already started ticking. Your code runs in a hosted container rather than on hardware you own, and for some teams that alone is disqualifying. And it is a paid product after the free hour, whereas Happy and Goose are free software. Cosyra is $29.99 a month after 1 hour free on signup; see pricing for the full picture. The closest cloud cousins are Claude Code on the web and GitHub Codespaces, both browser-first rather than native-app-first.

How do the remote-control coding tools compare?

The table below is the side-by-side we reach for when someone asks which of these to install. Five are relays (the agent runs on your machine); Cosyra is the cloud container (the agent runs on ours). Every third-party row is dated and grounded in a factsheet we re-verify on a schedule.

Two columns carry most of the decision. "Where compute runs" is the privacy and cost story: on your machine means your code and tokens stay put and the software is often free. "Machine must stay awake" is the reliability story: every relay needs a host you keep on, and the cloud container does not. The rest, covering platforms, price tier, and open source, sorts the relays among themselves once you have picked an architecture.

Relay or cloud: which should you pick?

One question settles the architecture, and a second sorts within it. First: can a computer you own stay powered and online whenever you want to code? If no, you want a cloud container, because every relay needs a live host. If yes, ask the second question: must your source code stay on machines you control? If yes, a relay with end-to-end encryption is the honest pick; if you do not mind a hosted box, either works and the cloud route saves you from babysitting a machine. The flowchart below is the same logic.

How do you code from your phone with no machine to keep awake?

The cloud route is the one with no host to manage, so the setup is short. Install the app, let the container provision, add your API key, and start an agent. The terminal session below is what the first two minutes look like once the container is up — no desktop client, no tunnel, no port forward.

That is the loop, and the difference from a relay is what you did not do: no laptop left running at home, no desktop client installed, no Cloudflare tunnel, no checking whether the host is awake before you start. Because the container hibernates instead of resetting, you can lock the phone, get off the bus, and reopen to the same session on a different device. For the agent-specific walkthroughs, see Claude Code on phone and Codex CLI on phone.

What does a real phone coding session look like?

Short and supervised, done in a gap between other things. We open the app on a train, on a couch, in a waiting room, type a prompt to the agent, watch it propose a change, and read the diff before accepting. The work that fits this well is broad: bug fixes, small features, dependency bumps, reviewing a branch, kicking off a test run. The work that fits badly is anything that wants a desktop IDE plugin or a heavy local GUI, which is true of the relay apps too, and none of these put a full IDE on the phone.

The opinion we will stand behind: the phone keyboard is fine for agent-driven coding, because you are mostly reading diffs and typing short prompts, not hand-writing hundreds of lines. Most people who say a phone is useless for coding are picturing themselves typing code character by character, which is not what agent-driven work is. Whether the agent runs on your own machine through a relay or on a cloud box, the phone-side experience is the same act: prompt, read, accept, repeat.

Which guide should you read next?

This pillar stays at the architecture level on purpose. For the specific tool or device, we go deeper here:

• Cosyra vs Happy Coder and Cosyra vs Goose, the two open-source relays, head to head with the cloud route.
• Cosyra vs Claude Code Remote Control, the first-party relay versus a hosted container.
• Cosyra vs Omnara, the free remote-control app — with a desktop client and Apple Watch companion — against a cloud container that keeps nothing awake.
• Claude Code on phone and Codex CLI on phone, the agent walkthroughs for the cloud route.
• SSH from your phone, the oldest relay of all — your own server, a terminal, and tmux.
• Mobile coding terminal: the complete guide, the device-agnostic view of the terminal layer underneath all of this.

Frequently asked questions