How to Run AI Coding Agents on Your Phone

You can run Claude Code, Codex CLI, OpenCode, and Gemini CLI on your phone right now. Install Cosyra for iOS or Cosyra for Android, add your API keys, and you get a full Ubuntu container with all four agents pre-installed. No SSH setup, no always-on laptop, no browser tabs that time out. We built this because we kept wanting to check on agent work from the couch or approve a Claude Code permission request while walking the dog, and the existing options were either fragile SSH chains or monitor-only dashboards.

This guide covers why phone-based AI coding works, how each agent runs on mobile, three approaches you can take, and step-by-step setup on Cosyra. If you want the broader context first (any terminal, not just AI agents), the mobile coding terminal pillar sets the stage, and if you are weighing which of these agents to run, AI coding agent CLIs compared ranks all eight on license, cost, and mobile reach. 1 hour free on signup, no credit card.

What do the four mobile AI coding agents actually do?

Claude Code is best for deep codebase understanding and multi-file refactors; Codex CLI is best for quick edits and sandboxed execution; OpenCode is the open-source option that works with 75+ providers including local models; Gemini CLI brings Google's 1M-token context window for large codebase navigation. Cosyra pre-installs all four and all four use BYOK, so you pay Anthropic, OpenAI, or Google directly at their standard rates.

Cosyra pre-installs four AI coding agents. All of them use BYOK (bring your own key), you connect your own Anthropic, OpenAI, or Google API key and pay the provider directly at their standard rates. Cosyra itself charges for compute, not AI tokens.

You don't have to pick one. All four are installed in the same container. Start a task with Claude Code, try a different approach with Codex, use OpenCode with a local model for private code. The container state is shared , they all see the same files, same git history, same environment variables.

Why do AI coding agents change the phone equation?

Because you direct the agent in short prompts instead of typing code character by character, and agents do most of the typing for you. Your job shifts to review and approve, which is a reading-heavy, tap-heavy interaction that fits a 6-inch screen. The standard "phone keyboards are too small to code on" objection is fair only when you are the one writing every semicolon. Agents flip that ratio, and the phone becomes viable for real engineering work.

With Claude Code or Codex CLI, you describe what you want in plain language and the agent writes the code, runs the tests, and proposes a diff. Your job is to direct, review, and approve. That interaction model, short prompts in, long outputs out, works on a phone because you're reading more than you're typing.

We think the phone keyboard is fine for agent-driven coding. Most developers who disagree haven't tried it with a properly set up agent that handles the boilerplate. When you're on the train reviewing a PR, or between meetings approving a Codex refactor, the phone is not a compromise. It's the fastest path between "I have five minutes" and "that shipped."

The practical scenarios where this matters:

• Monitoring long-running agent tasks. Claude Code working through a large refactor takes 10-20 minutes. You don't need to stare at a laptop for that. Check in from your phone, approve the permission it's waiting on, and get back to your coffee.
• Quick fixes and reviews. A bug report comes in while you're away from your desk. Open the terminal, point Claude Code at the issue, review the fix, push. Done before you get home.
• Learning and experimentation. You read about a new library on the bus. Spin up a scratch project, ask the agent to scaffold it, poke around. The container persists, pick it up later on any device.

What are the three approaches to AI agents on your phone?

The three approaches are SSH from your phone to a desktop or VPS that runs the agent, installing the agent directly in a local terminal app like Termux, or running the agent in a mobile cloud terminal like Cosyra. Each has a real trade-off: SSH needs an always-on machine, Termux runs the agents via npm but hits a glibc-vs-Bionic libc gap for the native installers plus Android's phantom process killer (we break down that specific tradeoff in our Cosyra vs Termux comparison, and we rank the three honest paths to a real Debian/Ubuntu userland on Android (proot-distro, UserLAnd, and a cloud x86_64 container) in Linux Container on Android: 3 Real Options in 2026), and a cloud terminal costs a subscription but removes both problems.

1. SSH from phone to a desktop or VPS

The classic approach: run the agent on a "real" machine and SSH in from Blink Shell (iOS) or Termux (Android). Add Tailscale for networking and tmux for session persistence.

• Upside: Full control, use your existing dev machine, free if you already have the hardware.
• Downside: Requires an always-on machine. If your MacBook sleeps or your VPS reboots, the session dies. Setup takes 15-30 minutes per machine. You're debugging networking instead of writing code.

This is the approach described in most existing guides, SSH + Tailscale + tmux + Blink. It works if you have a reliable server. We used it ourselves for months before building Cosyra.

2. Browser-based cloud IDEs on mobile

GitHub Codespaces and code-server run VS Code in a mobile browser. You get a full editor, terminal, and extensions.

• Upside: VS Code ecosystem, extensions, GitHub-native integration.
• Downside: Browser UIs on mobile are painful, tiny targets, constant mis-taps, Safari's address bar eating screen space. The GitHub Free personal-account quota is 120 core-hours per month, which is 60 hours of runtime on a 2-core codespace (verified 2026-05-19 in GitHub's billing docs). Orgs and enterprises get no free quota, only pay-as-you-go from $0.18/hr for 2-core. Either way, the browser experience on a phone is optimized for desktop screens, not 6-inch touch displays.

3. Native cloud terminal with agents pre-installed

This is what we built with Cosyra. A native iOS and Android app that connects to a persistent Ubuntu container in the cloud. Claude Code, Codex CLI, OpenCode, and Gemini CLI are pre-installed. The terminal UI is designed for touch, with larger tap targets, swipe gestures, and a toolbar row for common keys.

• Upside: Zero setup for agents, native mobile experience, container persists across sessions (30 GB storage, session hibernation).
• Downside: Paid service ($29.99/month after free trial). Requires internet, no offline mode. You're running on our infrastructure, not your own.

How do you set up each AI agent on Cosyra?

You set up each agent in about two minutes: install Cosyra, open the app to land in the Ubuntu container, paste the corresponding provider API key into your shell, and run the agent's command (claude, codex, opencode, or gemini). All four are already on PATH. You pay the provider directly at their standard token rates; Cosyra never touches your AI billing.

The setup takes about two minutes per agent. Here's what it looks like.

Step 1: Install and start a container

Download Cosyra from the App Store or Google Play. Create an account (email or Sign in with Apple/Google). The app provisions your Ubuntu container on first launch. In our testing, that took about 15 seconds.

Step 2: Add your API keys

Each agent needs its own API key. In the Cosyra terminal, set them as environment variables. They persist across sessions.

For OpenCode, configure your provider in its config file, it supports 75+ LLM providers including local models. See the OpenCode docs for provider setup.

Step 3: Clone a repo and start coding

That's it. The container has git, node, python, go, rust, and common dev tools pre-installed. Your files persist on 30 GB of storage. If the app goes to background, session hibernation keeps your work intact.

Step 4: Switch between agents in the same container

All four agents are regular CLI tools. Switch between them like you would on a laptop:

What does a real mobile AI coding session look like?

A real session looks like six minutes of phone time: open the app, the container is already running, cd into the project, run claude with a short "review the diff between main and this branch" prompt, read the structured review, approve the PR, and push a small fix. The agent does the heavy reading; you make the decisions. Below is a PR review we ran on the 8:40 AM train before standup at 9.

Here's a scenario we run through regularly. It's 8:40 AM, you're on the train to the office, and a teammate tagged you on a PR that needs review before standup at 9.

1. Open Cosyra on your phone. Your container is still running from yesterday.
2. cd into the project. git fetch && git checkout pr-branch.
3. Run claude and ask: "Review the diff between main and this branch. Focus on security issues and missing tests."
4. Claude Code reads every changed file, checks the test coverage, and gives you a structured review with specific line references.
5. You read the review on your phone, approve the PR with a comment referencing Claude's findings, and push a small fix Claude suggested.
6. You walk into standup at 9 having already reviewed and improved the PR. Total phone time: 6 minutes.

The agent did the heavy reading. You made the decisions. That's why the phone works, it's a command interface for an agent that does the actual typing.

What are the best tips for phone-based AI agent work?

The best tips are: let the agent do the typing with short descriptive prompts, bring an external keyboard only for sessions longer than twenty minutes, use tmux for split panes, commit and push from inside the container instead of switching devices, and rely on session hibernation so you can come back hours later with everything intact. Those five habits cut phone session friction more than any app setting.

• Use agent-first workflows. Don't try to write code character by character. Describe what you want in a sentence, let the agent generate, review the output. This is faster on a phone than manual editing on a laptop for many tasks.
• External keyboard is optional. For sessions longer than 20 minutes, a Bluetooth keyboard helps. For quick reviews and approvals, the phone keyboard is enough. We've shipped production fixes from the waiting room at the dentist with just a thumb.
• Use tmux inside the container if you want multiple terminal panes. Cosyra's container includes tmux pre-installed. tmux new -s work gives you split panes, one for the agent, one for watching test output.
• Commit and push from the container. The container has full git support. Connect your GitHub account via the Cosyra app (see Docs) and push directly.
• Session hibernation saves your state. Close the app, come back hours later, and your terminal session, running processes, and files are exactly where you left them. No reconnecting, no tmux reattach.
• Pick the right language runtime for your phone. Agents drive code, but they execute against whatever language your project uses. Our code Python on your phone breakdown covers exactly which pip packages install on a-Shell, iSH, Termux, and a cloud Ubuntu container, including the C-extension cliff that kills numpy and pandas on iOS.

How does mobile agent work on Cosyra compare to other mobile coding options?

Cosyra wins on pre-installed AI agents, native iOS and Android apps, persistent cloud compute that does not depend on your laptop being awake, and zero Android phantom-process-killer risk. SSH plus Blink or Termux is cheaper if you already run a server; Codespaces is deeper on GitHub integration; Termux is free but exposes you to ABI issues and phantom process killer on Android 12+. The table below lays them side by side.

Each option wins somewhere. Termux is unbeatable on price and offline capability. Codespaces is free and integrates with GitHub. SSH gives you maximum control. Our edge is that the agents work out of the box on a phone: no server maintenance, no browser compromises, no ARM compatibility issues.

Related guides: go deeper on the agent, device, or alternative you came for

Each guide below answers one specific phrasing of "I want to code from my phone." Pick the one closest to what you already have on your home screen.

One guide per agent

• Claude Code on your phone: Anthropic's agent, deep codebase understanding, multi-file refactors. On an iPhone specifically, Claude Code on iPhone walks the three real paths and why iOS has no Termux.
• Codex CLI on your phone: OpenAI's agent, sandboxed execution, fast iterative edits. On Android specifically, Codex CLI on Android covers the four install paths and why the official binary stopped working in Termux after v0.5.0. On an iPhone, Codex CLI on iPhone covers the four paths there, including the official ChatGPT-app remote.
• Gemini CLI on your phone: Google's terminal agent; the free OAuth tier ended 2026-06-18, so it now runs BYOK on a paid AI Studio or Vertex key. On an iPhone, see Gemini CLI on iPhone; on a tablet, Gemini CLI on iPad.
• OpenCode on your phone: open-source agent that pairs with any provider key you already have.
• Kilo CLI on your phone: the MIT-licensed agent built on OpenCode (so it shares the runtime we already run in production); a one-command npm install -g @kilocode/cli add on top of the four agents above.
• Crush on your phone: Charm's source-available terminal agent (FSL-1.1-MIT, 20+ providers); add it to a container in one command since it ships no mobile app, or see how the local binary stacks up against a cloud box in Cosyra vs Crush.
• Qwen Code on your phone: Alibaba's open-weight Gemini CLI fork (Node 22+, BYOK).
• Aider on your phone: the open-source Python pair-programmer with a git-native auto-commit loop. Covers the aarch64 wheel gotcha that breaks plain pip install on Termux.
• goose on your phone: the Apache-2.0 Linux Foundation agent whose free iOS app is a remote control, not an on-phone agent; how to run the actual agent in a cloud container instead of on a home server.
• Cline on your phone: the 62.9k-star Apache-2.0 agent with no mobile client of any kind; how to run its CLI (npm i -g cline) in a cloud container instead of keeping a laptop open. On a tablet, the gap runs two walls deep: Cline on an iPad covers why neither Cline nor the editors it embeds in ship an iPadOS build. Cosyra vs Cline weighs the editor-embedded agent against a real mobile terminal, feature by feature.
• GitHub Copilot CLI on your phone: GitHub's terminal agent is desktop-only, and the Termux install fails because its pty.node native module has no Android arm64 prebuild; GitHub's mobile feature only remote-controls a session running on a desktop that has to stay awake. For the head-to-head, see Cosyra vs GitHub Copilot CLI.
• Amazon Q CLI on your phone: AWS's terminal agent (q chat) is macOS- and Linux-only with no mobile app, and it authenticates through an AWS Builder ID rather than a provider key you bring. The open-source repo is now in maintenance mode (critical fixes only; successor is the closed-source Kiro CLI), but q still installs and runs — how to reach it from a phone in a cloud Ubuntu container. For the head-to-head, see Cosyra vs Amazon Q Developer CLI.
• Cursor CLI on your phone: Anysphere's terminal agent installs as a native binary from curl https://cursor.com/install -fsS | bash rather than npm, invokes as agent (not cursor-agent, the legacy community name), and bills through a Cursor subscription instead of a provider key you bring. There is no iOS or Android build; how to reach it from a phone in a cloud Ubuntu container.
• Jules on your phone: Google's async agent runs in a cloud VM and opens a pull request, so you can trigger and approve from a phone browser or a jules GitHub label, but it gives you no interactive terminal; how to keep a real shell alongside it. Cosyra vs Jules weighs that async-pull-request model against a live cloud shell.
• OpenHands on your phone: OpenHands is the open-source (MIT at its core) autonomous agent that picks up a whole task and runs it unattended in its own sandbox; there is no native app, but you can drive OpenHands Cloud from a phone browser. The guide walks the three real paths, and Cosyra vs OpenHands weighs that hands-off canvas against an interactive terminal you can drive from a phone when the agent stalls.
• Warp Terminal on your phone: the agentic terminal ships Mac, Linux, and Windows builds but no mobile app; how to get the same agent-in-a-shell setup from iOS or Android. For the head-to-head, see Cosyra vs Warp.
• Zed editor on your phone: Zed runs Claude Agent, Codex CLI, and Gemini CLI over ACP at the desk but ships no iOS, iPadOS, or Android build; how to run those same agent CLIs from a phone.
• Windsurf on your phone: Windsurf, renamed Devin Desktop in June 2026, runs Codex, Claude Agent, and OpenCode over ACP at the desk but ships no mobile build; how to run those same agent CLIs from a phone.
• Devin Desktop on your phone: after the 2026-06-02 Windsurf rename, the editor is still desktop-only, so you run its Codex, Claude, and OpenCode agents from a cloud container instead.
• AgentsRoom on your phone: AgentsRoom's iOS and Android apps remote-control agents running on your own always-on desktop across seven providers; how to run the same agent CLIs from a cloud container with no personal machine left awake.
• Kiro on your phone: Kiro, AWS's spec-driven agentic IDE, is built on Anthropic's Claude Agent SDK but ships no iOS, iPadOS, or Android app; how to run the same Claude agent from a phone.
• Continue's cn CLI on your phone: Continue was acquired by Cursor in mid-June 2026 and development on its main branch stopped (last commit 2026-06-19), but the Apache-2.0 cn terminal agent still installs and runs on your own API keys; how to run it from a phone inside a cloud container while the hosted Continue Hub winds down.
• Claude Code Skills on your phone: personal SKILL.md files in the persistent home directory, shared across iPhone, Android, and web.
• MCP servers on your phone: adding Model Context Protocol servers with claude mcp add, and why a local stdio server survives in a cloud container but gets reaped in Termux.

Picking between agents

• Best way to run AI coding agents on a phone: the decision guide for the three setups — on-device, relay to your own machine, or a hosted cloud box — before you pick an agent.
• Relay apps vs cloud containers: the architecture choice underneath the apps — nine relay apps (Happy Coder, Omnara, AgentsRoom, Moshi, and more) that run the agent on a machine you own, versus a hosted container that runs it for you.
• Claude Code vs Codex CLI on phone: approval defaults, token economics, when each wins.
• Claude Code vs OpenCode on phone: MIT vs custom license, multi-provider vs Anthropic-tuned surface.
• Claude Code vs Gemini CLI on phone: token economics, and what the Antigravity CLI transition on 2026-06-18 changed for Gemini users.
• AI pair programmer on your phone: three working paradigms (remote-control desktop, cloud container, local Termux) compared honestly.
• Vibe coding on your phone: the developer version of that term (agents editing your real repo) vs the no-code app builders.

By device

• Claude Code on iPad: Magic Keyboard, Split View, Stage Manager workflows.
• Claude Code on Android: Pixel, Galaxy, foldables, Samsung DeX. Covers the Bionic-vs-glibc ABI gap and Android 12+ phantom process killer.
• Claude Code on a Chromebook: local Linux container (Crostini) vs a browser tab.

Compared to other tools and cloud IDEs

• Firebase Studio on iPad: Google's hosted cloud IDE in Safari on a tablet.
• Jules on iPad: Google's async agent on a tablet: the responsive web app and a keyboard make trigger and review comfortable, but there is still no terminal and no iPadOS app.
• Firebase Studio on phone: same workspace, phone-sized screen (train, couch, waiting room).
• Gitpod (now Ona) on iPad: hosted IDE with no native app, post-2025 rebrand.
• Cosyra vs Moshi: Moshi drives Claude Code on a Mac or VPS you own from an iOS app with Apple Watch approvals.
• Cosyra vs Cursor: the desktop-first AI IDE with a web/mobile PWA for monitoring cloud agents. For the phone-specific walkthrough of what that PWA can and can't do, see Cursor on your phone. Its terminal agent is a separate product with its own install path, covered in Cursor CLI on your phone.
• Cosyra vs Windsurf: Windsurf, the other VS Code fork, was renamed Devin Desktop on 2026-06-02 and stays desktop-only (macOS/Windows/Linux) with no mobile build. Read this if you ran Windsurf at your desk and want its Codex, Claude, and OpenCode agents from a phone.
• Cosyra vs Catnip: Catnip is an open-source iOS app (W&B / Apache 2.0) that wires Claude Code into a GitHub Codespace. Read this if you already have Codespaces seats and want a phone-shaped UI on top of them.
• Cosyra vs Happy Coder: Happy Coder is an open-source (MIT) iOS/Android app that relays your phone to a Claude Code session running on your own machine, with realtime voice. Read this if you keep a desktop awake and want a polished mobile front end rather than hosted compute.
• Cosyra vs Kiro: Kiro (AWS) is a spec-driven agentic IDE built on Anthropic's Claude Agent SDK. Desktop-only (macOS/Windows/Linux) with a paid browser dashboard for autonomous PRs and no mobile app. Read this if you run Kiro at your desk and want the same Claude lineage from your phone.
• Cosyra vs Aider: Aider is the open-source, model-agnostic CLI pair-programmer with no mobile app; Cosyra is the cloud Ubuntu terminal you run it inside. Read this if you already live in Aider at your desk and want the same git-native loop from a phone.
• VS Code on your phone: vscode.dev opens in the mobile browser but has no terminal and can't run your code; the guide covers what edits there and how to get a real shell instead.
• VS Code on iPad: the same vscode.dev limits in Safari even with a Magic Keyboard, and the cloud-container route around them. For the head-to-head, see Cosyra vs VS Code for the Web.

Frequently asked questions

Get started now. Download Cosyra and run your first AI agent from your phone in two minutes.

Download on App Store Get it on Google Play

Scan to install on your phone iPhone or Android

See pricing