Android Studio now runs Claude, Codex and 40-plus AI agents via ACP

Android Studio opens up to outside coding agents

At Android Dev Summit on October 28, Google announced that Android Studio now supports more than 40 ACP-compatible coding agents directly inside the IDE, through a feature called Bring Your Own Agent (BYOA) built on the open Agent Client Protocol (ACP). According to the Android Developers Blog, the capability ships stable in the Android Studio Rabbit 2 release. As a dev.to write-up of the announcement points out, this targets developers who have been running Claude Code or Codex in a terminal window next to the IDE just to reach the emulator, Logcat and previews.

The named agents include Claude Agent, OpenAI's Codex, Junie and Google's own Antigravity, and Google says any ACP-compliant agent can plug in. The marketplace sits under Settings > Tools > AI > Agents, where developers can install several agents and switch between them; The Decoder reports that they can run simultaneously. The pitch is that agents gain the IDE's project awareness: understanding of project structure and build, symbol resolution, Compose Preview rendering, the emulator and Logcat.

One protocol, many agents

The plumbing is the Agent Client Protocol, an open specification for how editors and coding agents communicate. Per the ACP documentation, local agents run as subprocesses of the editor and exchange messages over JSON-RPC on stdio. That design explains how a single integration covers dozens of agents: Android Studio implements the protocol once, and any agent that speaks ACP, natively or through an adapter, can connect.

Rabbit 2 also adds Change Review. A review session can start from an agent's pending changes or the Git changes window, and Android Studio generates a report of potential issues; it is also reachable through a /review command in the agent tool window.

Android CLI skips the IDE entirely

The quieter half of the announcement: Google's Android CLI now installs as an official plugin for Claude Code, Codex and Antigravity, and it does not require Android Studio to be open at all. Google positions it for any terminal, any agent harness and CI pipelines, letting agents build, run and debug Android apps autonomously.

Installation goes through each agent's plugin marketplace, for example claude plugin marketplace add android/android-cli-plugin followed by a plugin install command; the dev.to write-up notes the Codex variant uses add rather than install in the second step, and Antigravity has its own form. The plugin repository is public under Apache-2.0, but the plugin itself is only a skill that teaches the agent to drive the CLI — Android CLI remains a separate download, initialized with android update and android init. A --no-metrics flag disables data collection, and a ~/.androidrc file applies flags automatically, though the docs do not confirm the metrics flag works there.

The capabilities include creating projects, managing virtual devices, inspecting live UI layouts, running Journeys and searching the official Android Knowledge Base. Two additions stand out: android preview renders Compose Previews without an emulator or the IDE, and android device remote can provision a real physical device from a secure Google device lab.

Read the speed numbers carefully

Google's internal testing claims agents using android preview completed UI tasks 77% faster while using up to 3x fewer tokens, and that on a LeakCanary memory-leak task the most capable models ran almost 2x faster with up to 25% fewer tokens — an A/B test The Decoder says used Claude Opus 5.5. As the dev.to analysis notes, these are vendor tests on vendor-picked tasks with no published task set. The direction is plausible, since rendering a preview should cost fewer turns than booting an emulator and screenshotting it, but teams should measure on their own apps.

Privacy and permission caveats

The privacy guarantees in the release post are written for Gemini Enterprise: source code and workspace context are not retained or used to train global foundation models, with audit logging and central admin policy on top. Google does not say those terms extend to Claude Agent or Codex, so a third-party agent inside Android Studio still operates under its own vendor's data terms. Sign-in is per agent, each with its own account and usage limits. The dev.to write-up also recommends gating device provisioning behind an approval prompt, the way teams already treat git push, since a tool that can allocate real hardware deserves the same scrutiny.

Why it matters

Mobile tooling is converging on the same pattern seen in general-purpose editors: an open protocol, rather than per-vendor integrations, as the connective tissue between IDEs and AI agents. With ACP as the interface, Android Studio can host Claude, Codex or whatever ships next without new plumbing, and other editors could do the same. Just as notable, the Android CLI plugin moves agent support outside the IDE altogether, into terminals and CI pipelines — a signal that Google now treats autonomous agents as a primary audience for Android tooling rather than a novelty. The open questions are whether data terms for non-Google agents catch up, and how much autonomy teams will grant tools that can provision physical devices.