# Android Studio Rabbit 2: Run Any AI Agent in Your IDE > Source: > Published: 2026-11-05 22:08:38+00:00 Android Studio Rabbit 2 went stable on October 28, and the headline change is simple: Google is no longer the gatekeeper of which AI coding agent you use. Through a new “Bring Your Own Agent” feature built on the open [Agent Client Protocol (ACP)](https://www.morphllm.com/agent-client-protocol), 40+ external agents — including Anthropic’s Claude Agent and OpenAI’s Codex — now get the same deep IDE access Gemini kept to itself. We’re talking full project graph, build diagnostics, Jetpack Compose previews, Logcat, and emulator control. Not a chat window. Real tools. ## What agents actually get The difference between AI-in-your-IDE and AI-with-your-IDE is tool access. BYOA bridges that gap. When you connect an ACP-compatible agent, it receives your complete project graph, build configuration, symbol resolution, and platform details. It can render Compose previews, read Logcat, trigger builds, run shell commands, and control the emulator directly. It can even spawn subagents for specialized tasks — code review, testing, security audit — and maintain long-running sessions across your project. Granular permissions mean the agent acts on routine work autonomously and pauses for your approval before riskier actions. That’s a meaningful distinction. You’re not handing your codebase to an agent running blind. ## ACP: the LSP for coding agents The Agent Client Protocol is worth understanding beyond the Android context. ACP is an open JSON-RPC standard — Apache license, created by Zed Industries in August 2025, now adopted by JetBrains, Google, and GitHub. It reached stable v1 (schema 1.24.1) in October 2026 with 25+ agents in its registry. The analogy that holds up: ACP is to AI coding agents what LSP was to language servers. Before LSP, every editor needed a custom integration for every language. After LSP, language servers plugged into any editor that spoke the protocol. ACP does the same for agents — implement once, work in any IDE that speaks ACP. ACP and MCP are complementary, not competing. ACP connects your editor to the agent (bidirectional, session-managed, streams tokens in real time). MCP connects the agent to external tools and data. They wire up in a single handshake: when an ACP session starts, it declares the MCP servers the agent should connect to. Full stack, one initialization. [CircleCI’s breakdown of ACP vs MCP](https://circleci.com/blog/acp-vs-mcp-whats-the-difference-for-agentic-coding/) is a good reference if you want the protocol-level detail. ## Setup: three steps Getting a third-party agent running in Android Studio takes about five minutes: 1. Update to Rabbit 2 via **Help > Check for Updates** and confirm you’re on a supported Android Gradle Plugin version (7.1 through 9.5) 2. Open **Settings > Tools > AI > Agents** , browse the ACP registry, and select your agent 3. Authenticate with your provider credentials or API key, then start with a read-oriented request to verify the connection Google’s [official BYOA announcement](https://developer.android.com/blog/posts/build-your-way-use-any-ai-agent-of-your-choice-in-android-studio) covers the full feature list. The [Rabbit 2 release notes](https://developer.android.com/studio/releases) include the AGP compatibility matrix and full changelog. ## CLI plugin: agents without the IDE If you work primarily in Claude Code or Codex rather than Android Studio, there’s an official Android CLI plugin that brings Android-specific capabilities to your terminal agent. Install it in one command: ``` # Claude Code claude plugin marketplace add android/android-cli-plugin # OpenAI Codex codex plugin marketplace add android/android-cli-plugin ``` Once installed, the plugin lets agents render Compose previews without an emulator, manage virtual devices, run Journeys, integrate LeakCanary and Firebase Crashlytics, and stream from remote physical devices. That last capability is useful for hardware-specific bugs that emulators don’t reproduce. ## The numbers — with the caveat Google reports that agents using the terminal-based Compose preview command completed UI tasks 77% faster while using up to 3x fewer tokens. Memory leak detection with LeakCanary integration ran nearly twice as fast with about 25% fewer tokens. These are vendor-conducted tests on undisclosed task sets. Treat them as directional, not gospel. Android Bench 2.0 — Google’s new benchmark for multi-day engineering tasks — should produce more independently verifiable comparisons as community results accumulate. ## The honest state of “stable” Google shipped Rabbit 2 to the stable channel, but community reports tell a different story. Developers on Kotlin Multiplatform projects describe Gradle sync failures, garbage collection problems, and IDEs they had to force-quit multiple times a day. Code analysis is reportedly missing during commit flows. This doesn’t mean you shouldn’t use it. It means you should start in a non-production repository, run your full test suite after any agent-generated change, and treat “stable” as “ready to evaluate” rather than “battle-hardened.” ## Security: your code, their policies Connecting a third-party agent to your codebase means your source files, build logs, and potentially emulator output travel under that vendor’s data retention policies. Gemini Enterprise has audit logging and admin controls. Claude Agent and Codex operate under Anthropic’s and OpenAI’s terms respectively — verify those terms before connecting to a sensitive repository. Use least-privilege API credentials and don’t store keys in tracked files. Review agent-suggested patches for new permissions, exported components, or dependency additions before merging. These aren’t hypothetical risks; agents with tool access can make changes you didn’t explicitly authorize. ## Where to start If you already use Claude Code or Codex, the CLI plugin is the lowest-friction entry point. Install it, ask the agent to render a Compose preview, run a UI task, and measure the actual improvement on your project. The 77% figure may hold — or it may not. Your codebase is not Google’s benchmark. If you’re inside Android Studio, install Rabbit 2, connect one agent through the registry on a throwaway branch, and run a focused UI task. Compare the output to what you’d write manually, check token usage, and form your own view on whether BYOA earns a place in your workflow. Either way, ACP is worth tracking regardless of your Android interest. It’s the interoperability layer that prevents IDE lock-in for coding agents, and it’s already running in Zed, JetBrains IDEs, and now Android Studio. The agents you wire up today via ACP will work in whatever editor you’re using in two years.