Android Studio Now Runs Claude and Codex Through ACP. Android CLI Doesn't Need the IDE

If you build Android apps with Claude Code or Codex, you have probably been running two windows: the agent in a terminal, and Android Studio next to it for the emulator, Logcat and previews. On October 28, at Android Dev Summit, Google moved both of those pieces toward the agent you already use.

First, Android Studio itself now hosts outside agents. In Google's words, "Android Studio now supports over 40 ACP-compatible coding agents directly inside Android Studio through Bring Your Own Agent and the open Agent Client Protocol (ACP)." The same post says the feature "is now stable in today's release of Android Studio Rabbit 2."

Second, and easier to miss: Google's Android CLI now installs as an official plugin for Claude Code, Codex and Antigravity. That one doesn't need the IDE open at all.

Bring Your Own Agent (BYOA) is a marketplace inside the IDE. Google's instructions: "open Settings > Tools > AI > Agents to browse and install agents from the marketplace. You can install more than one and switch between them as you work" ( Android Developers Blog ).

The Rabbit 2 release post names four of the 40-plus agents: Claude Agent, Codex from OpenAI, Junie and Google Antigravity, and says any "ACP compliant agent" works. The Decoder reports they can run at the same time.

The pitch is that the agent gets the IDE's knowledge of your project. Google lists "a deep understanding of your project structure and build, symbol resolution, Compose Preview rendering, the Emulator, Logcat, and more" ( Android Developers Blog ).

The plumbing is the Agent Client Protocol, which "standardizes communication between code editors/IDEs and coding agents." Local agents run as subprocesses of the editor and talk "via JSON-RPC over stdio," per the ACP site. That is why one integration can cover 40 agents: Android Studio speaks the protocol, and any agent that speaks ACP, natively or through an adapter, can plug in.

Rabbit 2 also adds Change Review. You "start a review session from pending agent changes or the Git changes window, and Android Studio generates a report of potential issues," and you can call it with a /review command from the agent tool window ( Android Developers Blog ).

One protocol, many agents: Android Studio speaks ACP, and each agent plugs in through it.

Google describes Android CLI as working "in any terminal, with any agent harness, and in your CI pipelines, so your agents can build, run, and debug Android apps autonomously" ( Android Developers Blog ).

The plugin docs give these commands. For Claude Code, from your terminal:

claude plugin marketplace add android/android-cli-plugin claude plugin install android-cli@android-cli-plugin Enter fullscreen mode Exit fullscreen mode For Codex:

codex plugin marketplace add android/android-cli-plugin codex plugin add android-cli@android-cli-plugin Enter fullscreen mode Exit fullscreen mode Note the Codex verb is add, not install, in the second line. Antigravity gets agy plugin install android-cli-plugin@antigravity-plugins-official.

The plugin lives in a public android/android-cli-plugin repo under Apache-2.0. Its README says the repo "holds the plugin only," which matters: the plugin is a skill that teaches the agent how to use Android CLI, and the CLI itself is a separate install. You still install Android CLI from the downloads page, then, per the CLI docs:

android update android init Enter fullscreen mode Exit fullscreen mode android init installs the android-cli skill. The docs also say you can turn off data collection with the --no-metrics flag. If your agent runs in CI, decide that on purpose and pass the flag where the pipeline calls the tool. They also describe a ~/.androidrc file that applies flags automatically, one per line; they don't say whether --no-metrics works there, so check before you rely on it.

What the agent gets, per the plugin docs: it can "create projects, manage virtual devices, inspect live UI layouts, run Journeys, and search the official Android Knowledge Base." The summit post adds android preview, which "renders your Compose Previews without needing an emulator or the IDE," and android device remote, which can "provision a real physical device from a secure Google device lab."

The headline figure is this sentence: "In our internal testing, agents that used android preview completed UI tasks 77% faster while using up to 3x fewer tokens" ( Android Developers Blog ). For memory leaks, Google says that with Android CLI's LeakCanary tool "the most capable models ran this process almost 2x faster while using up to 25% lesser tokens." The Decoder reports that A/B test used Claude Opus 5.5.

Both are vendor tests on vendor-picked tasks, and the post doesn't publish the task set. The direction is believable. An agent that can render a preview instead of booting an emulator and screenshotting it should spend fewer turns. Measure the gain on your own app before you quote it.

android preview renders Compose Previews from the terminal, so the agent can check UI without booting an emulator.

The privacy guarantees in the release post are written for Gemini Enterprise. The Rabbit 2 post says that for Gemini Enterprise, source code and workspace context are never retained or used to train global foundation models, and it adds audit logging and central admin policy. It doesn't say those terms extend to Claude Agent or Codex. A third-party agent inside Android Studio is still that vendor's agent, on that vendor's terms. Check your Anthropic or OpenAI plan's data terms before you turn either one on.

Sign-in is per agent. Google's setup step is to "authenticate" ( Rabbit 2 post ), with no detail on what each agent requires. Expect each one to use its own account and its own usage limits.

Remote devices and preview rendering give an agent more reach. If a tool can provision a physical device in Google's lab, put it behind an approval prompt, the same as git push. Keep your agent's permission mode where it is now, and let it ask before it runs android device remote.

I haven't run Rabbit 2 or the plugin myself. Everything above is from Google's October 28 posts and its Android CLI docs.

I build Hermes IDE, so weigh this with that in mind. Hermes runs Claude Code and Codex in a real terminal, "exactly as it would anywhere else, so a vendor's new commands and modes work in Hermes the day they ship" ( Hermes README ). The plugin installs into the agent, not the editor, so the same two claude plugin or codex plugin lines are what you'd type in a Hermes session.

What Hermes adds is a separate hermes/<task> branch and worktree per task, so you could give Claude Code and Codex the same Android ticket and compare. It does not render Compose Previews or host an emulator. For those, Android Studio's own BYOA panel is the tool built for the job. Hermes is free, on macOS, Windows and Linux, and not affiliated with Google, Anthropic or OpenAI.

Codex on orders-types and Claude Code on orders-api, side by side in Hermes 2.3.4, each in its own hermes/ worktree, with status, context use, cost and branch in the sidebar.

So the question for Android teams is which surface you want the agent in. Do you want it inside Android Studio, next to Logcat, or in a terminal where the CLI plugin gives it the same tools? Tell me which one you're trying first.

An agent that can build, preview and run your Android app from a terminal is an agent that can do it in CI too. Ship It is about that whole stretch between localhost and production for a small team: hosting, CI/CD, auth, payments and monitoring, so the pipeline the agent lands in is one you trust. If you want to try the CLI plugin in a real session, Hermes IDE is free and runs Claude Code and Codex side by side.