Best AI Coding Agent CLIs for Your Phone (2026)

The best AI coding agent CLI for your phone is the one whose model you trust and whose license you can live with. The agents themselves have converged on the same terminal shape. Claude Code, Codex CLI, OpenCode, and Gemini CLI all run as a terminal program you drive with plain English; Cursor CLI, Amp, Crush, and Aider round out the field. This guide compares all eight on the four things that actually decide it from a phone: license, cost model, install path, and how each one reaches a small screen. We pre-install the first four because switching between them is a one-word command.

We build Cosyra, so we have a side here, but the useful version of this comparison is the one that tells you which agent to run, not which company to like. We installed the four we ship into a clean Ubuntu 24.04 container on 2026-07-16 and read their versions back off the shell, so the numbers below are first-hand, not copied from a release page. If you want the setup mechanics for any single agent, the AI coding agents on mobile pillar has the step-by-step; this page is the choosing part underneath it.

The eight AI coding agent CLIs, compared

These eight cover the agents developers actually reach for in a terminal in 2026. They split cleanly on two axes: how open the code is, and how you pay for the model. Everything else (approval defaults, install path, mobile reach) falls out of those two. Every "as of" date is when we last verified that row first-hand against the tool's own docs, repo, or npm metadata.

The two Anthropic-and-OpenAI defaults: Claude Code and Codex CLI

Claude Code is the agent most people mean when they say "coding agent in the terminal." It is proprietary, from Anthropic, and ships on a fast cadence. We installed 2.1.197 on 2026-07-16, and it moves several point releases a week. You bring your own Anthropic key or drive it from a paid Claude plan. Its default is conservative: it asks before running commands that change your machine, which on a phone, where you cannot glance at a second monitor, is the behavior you want.

Codex CLI is OpenAI's answer, and it is the opposite trade in two ways. It is open source (Apache 2.0, the code is at openai/codex), and it defaults to looser approvals. It will run shell commands, write files, and run npm or pip inside the workspace without asking each time. That is faster and more dangerous. On a phone we run it with codex --ask-for-approval always when we are on a network we do not fully trust. We installed codex-cli 0.144.5 first-hand on 2026-07-16.

The open pair: OpenCode and Aider

OpenCode is the most open agent on this list: MIT-licensed, model-agnostic, and, at 186,323 GitHub stars on 2026-07-16, the one with the most momentum. It does not tie you to a single provider — point it at Anthropic, OpenAI, Google, or a local model. The one honest wrinkle for phone users: it still misdetects Termux as plain Linux (issue #21043, opened 2026-04-04, auto-closed by a staleness bot on 2026-07-03 without a fix), so on raw Android you fall back to a proot chroot. In a cloud container that problem never comes up because the container is plain Linux.

Aider is the elder statesman, an Apache-2.0 Python agent (aider-chat on PyPI, 0.86.2) that pioneered a lot of the git-native, edit-in-place workflow the newer agents copied. It is model-agnostic and works well with Gemini 2.5 Pro, DeepSeek, Claude, and o3-class models. The caveat is momentum: as of 2026-06-11 its main branch had gone ~20 days without a commit, so the release cadence has clearly slowed relative to Claude Code or OpenCode. If you value stability over the newest features, that is a feature, not a bug.

Gemini CLI and the free-tier that ended

Gemini CLI is Google's Apache-2.0 agent (@google/gemini-cli, installed first-hand at 0.50.0 on 2026-07-16). The thing to know before you pick it: the free first-party endpoint ended on 2026-06-18. The free OAuth login now returns 410 Gone, with no grace period. The binary still works, but you have to feed it a paid Gemini API key from Google AI Studio or Vertex AI. Google's official successor is Antigravity CLI (the av command), which did not ship at feature parity. Do not conflate "the free tier ended" with "the tool is dead." The toolchain is under active daily development.

The three you install yourself: Cursor CLI, Amp, Crush

Cursor CLI (the agent command) is Anysphere's headless runtime, still labelled beta. Two things make it different: it installs as a native binary via a curl script rather than from npm, so glibc and CPU arch matter the way they do for any compiled binary; and Cursor's mobile story is a desktop story: the iOS app launches and monitors cloud agents but gives you no terminal on the phone. If you live in the Cursor editor, the CLI is a natural extension. If your phone is the point, it is the wrong tool.

Amp, from Sourcegraph, is the one that breaks the BYOK pattern. You do not paste your own key. You buy Amp credits ($5 minimum) and Amp routes each request to a model it picks. Its @ampcode/cli package publishes daily and runs on macOS, Linux, and Windows via WSL. Its "Agents, Everywhere" mobile surface lets you watch and steer a running thread from a phone, but the agent is executing on a desktop-class machine, not on the phone. That is the honest framing: Amp on a phone means monitoring, not running. We break down the credit-billing trade-off, model routing, and where Amp beats a BYOK container in Cosyra vs Amp.

Crush, from Charmbracelet, is the most phone-friendly of the install-it-yourself three for one structural reason: it is a single Go binary with a prebuilt android-arm64 archive. A self-contained Go binary does not fight Termux's aarch64 userland the way a Node or Python agent does, so Crush is the one agent here you can genuinely run on raw Android without a chroot. Its license is source-available FSL-1.1-MIT (free to use, converts to MIT after two years), which is more permissive than proprietary but not OSI open source, worth naming so you do not call it "open source" and get corrected.

Where each agent genuinely wins

No hedging. Here is the one-line case for each, including the ones we do not ship:

• Claude Code wins on judgment and safe defaults. If you want the agent to ask before it runs something destructive, this is the one.
• Codex CLI wins on open source plus speed. It is Apache 2.0 and it does not stop to ask, which suits a trusted sandbox.
• OpenCode wins on openness and model choice. MIT, the biggest community, and no provider lock-in.
• Gemini CLI wins if you are already paying for Gemini and want Google's models in the terminal, if you accept the paid-key requirement first.
• Cursor CLI wins if you live in the Cursor IDE and want its agent in a terminal on your desktop. Not a phone tool.
• Amp wins if you would rather buy credits than manage keys and you want Sourcegraph's model routing to make the choice for you.
• Crush wins on raw Android (a single Go binary that runs in Termux without a chroot) and on a clean TUI.
• Aider wins on stability and a mature git-native workflow, if you do not need the newest features.

Here is the opinion a lot of these vendors would push back on: the agent you pick matters less than the model behind it and the box you run it in. They are all BYOK-ish terminal programs converging on the same edit-in-place, run-tests, approve-diffs loop. We ship four rather than betting on one because switching between claude, codex, opencode, and gemini is a single word at the prompt, and the differences between them shrink every month. Cursor and Amp, whose whole pitch is a differentiated managed experience, would tell you the opposite. We think they are selling the wrapper when the model is the product.

How each one actually reaches a phone

This is where the field splits hardest, and it is the axis desktop-centric blogs skip. Reviewing an agent's diff on the 7:52 train is a different problem from running it at a desk, and the agents were not all built for it:

• Terminal agents you reach remotely: Claude Code, Codex, OpenCode, Gemini, Aider, Crush. These have no phone app; "on a phone" means running them in a Linux environment you connect to. That is either Termux on Android (where the Node and Python ones fight the userland; see the Android container wall) or a cloud container that gives your phone a full shell.
• Desktop agents with a mobile remote: Cursor CLI and Amp. The agent runs on a desktop-class machine; the phone launches or monitors it. Useful if you already keep a machine awake, useless if the phone is your only always-on device.

We built Cosyra for the first group. Each user gets a private Ubuntu 24.04 container on Azure AKS with Claude Code, Codex CLI, OpenCode, and Gemini CLI already on PATH, 30 GB of storage that survives between sessions, and native iOS and Android apps. Because the container is x86_64 plain Linux, the same environment where all four installed cleanly in our 2026-07-16 test, so there is no ABI gap to fight. The honest trade-off: the container lives in the cloud, so no network means no terminal, the way a local tool like Termux keeps working on a plane. If offline is your hard requirement, run Crush in Termux and skip us.

Which should you pick?

Here is the framework we actually give people who ask:

• Pick by model loyalty first. If you trust one provider's models (Anthropic, OpenAI, or Google), start with that vendor's agent (Claude Code, Codex CLI, or Gemini CLI). The model does most of the work; the CLI is the harness.
• Pick OpenCode if you want to stay provider-neutral and auditable: MIT, model-agnostic, the largest community, and the safest bet against any one vendor changing terms.
• Pick Crush if you need raw Android with no cloud and no chroot. The Go binary is the only one here that runs cleanly in Termux.
• Pick Amp or Cursor CLI only if you want credits-over-keys (Amp) or you already live in the Cursor IDE, and accept that neither gives you a real terminal on the phone.

Try one you do not have to install first if the phone is your point. The four we pre-install are on PATH the second the container boots, so you can compare Claude Code against Codex against OpenCode against Gemini on the same repo, from the same phone, without four separate setups. If one of them wins for your work, you will know inside an afternoon. And if none of them do, the honest answer might be Crush in Termux or Cursor on your desktop, and we will have told you so.

Frequently asked questions

Compare all four on the same repo. Claude Code, Codex, OpenCode, and Gemini are on PATH the moment the container boots: no install, no arch fight, from your phone.

Download on App Store Get it on Google Play

Scan to install on your phone iPhone or Android

See pricing · AI agents on mobile pillar · Mobile terminal pillar