{"slug": "android-studio-rabbit-2-run-claude-and-codex-inside-the-ide", "title": "Android Studio Rabbit 2: Run Claude and Codex Inside the IDE", "summary": "Google released Android Studio Rabbit 2 as stable this week, adding a Bring Your Own Agent feature that runs 40+ coding agents — including Claude, Codex, and Google's Antigravity — inside the IDE via the open Agent Client Protocol. Google also added 30+ official Android skills for grounding agents in current APIs, and its internal testing found agents using the `android preview` command completed UI tasks 77% faster while using up to 3x fewer tokens. The release matters because ACP, JSON-RPC 2.0 over stdio, is also supported by JetBrains IDEs, Zed, Neovim, Emacs, and VS Code, positioning it as a universal adapter layer between editors and coding agents.", "body_md": "Google shipped Android Studio Rabbit 2 as stable this week, and the headline feature is a direct fix for a workflow that has quietly frustrated Android developers for years: running Claude or Codex in a separate terminal while constantly switching back to the IDE for context. The new Bring Your Own Agent (BYOA) feature lets you install and run 40+ coding agents — including Claude, Codex, and Google’s own Antigravity — directly inside Android Studio, all through an open standard called the Agent Client Protocol.\n\n## What BYOA Actually Gives You\n\nThe difference between running an agent in a terminal and running it inside Android Studio is context. A terminal agent guesses at your project structure, burns tokens on Android API documentation it may or may not have in training, and can’t see your Compose Preview or Logcat output. BYOA changes that.\n\nWhen you install an agent through **Settings > Tools > AI > Agents**, it gets full IDE-native access to your project graph, build diagnostics, symbol resolution, Compose Preview rendering, the Emulator, and Logcat. The agent isn’t just reading files — it understands your build the way Android Studio does. You can install multiple agents and switch between them mid-task.\n\nGoogle also added 30+ official Android skills to ground agents in current Android APIs. These aren’t generic programming tips; they target fast-moving areas where models often fall short: Compose performance optimizations, Camera APIs, security audits. When you run `android skills add jetpack-compose-performance`, you’re giving your agent project-aware guidance, not just documentation links.\n\n## The Agent Client Protocol: Bigger Than Android\n\nThe open standard underneath BYOA is worth understanding on its own terms. The [Agent Client Protocol (ACP)](https://agentclientprotocol.com/get-started/introduction) is JSON-RPC 2.0 over stdio — the same transport mechanism as the Language Server Protocol. The LSP analogy holds up well: before LSP, every editor needed its own TypeScript parser, Python linter, and Go formatter. ACP does the same thing for AI coding agents. One compliant agent works in any supporting editor, no custom glue code required.\n\nAndroid Studio isn’t the only IDE with ACP support. JetBrains IDEs, Zed, Neovim, Emacs, and VS Code all support it. ACP is becoming the universal adapter layer between editors and agents, and it’s worth watching even if Android development isn’t your focus.\n\nACP and MCP serve different layers: ACP handles editor-to-agent communication, MCP handles agent-to-tools communication. They complement each other — an ACP session can declare which MCP servers the agent should connect to. Android Studio uses both: BYOA (ACP) for running your agent inside the IDE, and AppFunctions (now in beta) for letting your app expose callable functions to agents via Android’s MCP protocol.\n\n## Android CLI: The Terminal-First Path\n\nIf you work CI-first or prefer the terminal, the IDE isn’t required. [Android CLI](https://developer.android.com/tools/agents/android-cli) gives any agent, terminal, or pipeline first-class access to the Android toolchain, including real physical devices via Google’s device lab.\n\nInstalling the official Android plugin for Claude Code takes two commands:\n\n```\nclaude plugin marketplace add android/android-cli-plugin\nclaude plugin install android-cli@android-cli-plugin\n```\n\nThe standout command is `android preview`, which renders Compose Previews without spinning up an emulator. Google’s internal testing shows agents using it completed UI tasks 77% faster while using up to 3x fewer tokens. Take those figures as directional rather than guaranteed — they’re vendor-tested on Google’s workloads — but the efficiency gain from skipping emulator boot time is real.\n\nMemory profiling gets the same treatment. `android profiler leakcanary` analyzes heap dumps on your machine rather than the device. In Google’s testing, the most capable models ran this process almost twice as fast with 25% fewer tokens compared to doing it without Android CLI.\n\n## Android Bench 2.0: Pick the Right Agent\n\nWith 40+ agents available, the natural question is which one to use. The [Android Bench 2.0 leaderboard](https://developer.android.com/bench) evaluates models on long-horizon tasks — the kind that take an Android engineer multiple days: upgrading dependencies, adding features, building an app from mocks. Scoring is continuous (functionality, visual fidelity, regression avoidance), not a simple pass/fail.\n\nCurrent standings: Claude Opus 5.5 leads at 32% on long-horizon task pass rate, followed by GPT-6 Astra at 28%. Even the top models cap out around 80% on cross-platform-to-Android porting. Android Bench is worth checking before committing an agent to a large migration — the differences between models on Android-specific tasks are more significant than general benchmarks suggest.\n\n## What’s Stable, What Isn’t\n\nBYOA, Android CLI, and the XR emulators are all stable in Rabbit 2. The XML-to-Compose Migration Assistant is Canary. A2UI (agent-generated Compose screens from declarative JSON) and the Lightbuild declarative build system are alpha — interesting for experimentation, not production. The on-device Gemini Nano 4 agent loop is also alpha and requires Android 14 and AICore enrollment.\n\nThe official [Android Dev Summit post](https://developer.android.com/blog/posts/android-dev-summit-26-developer-tools-performance) has the full breakdown by stability tier. Start with the stable features and check the Canary builds once the core BYOA workflow feels comfortable.", "url": "https://wpnews.pro/news/android-studio-rabbit-2-run-claude-and-codex-inside-the-ide", "canonical_source": "https://byteiota.com/android-studio-rabbit2-bring-your-own-agent/", "published_at": "2026-11-04 01:07:24+00:00", "updated_at": "2026-11-04 01:21:38.497330+00:00", "lang": "en", "topics": ["ai-agents", "agent-protocols", "ai-tools", "developer-tools", "ai-products"], "entities": ["Google", "Android Studio Rabbit 2", "Bring Your Own Agent", "Agent Client Protocol", "Claude", "Codex", "Antigravity", "Android CLI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/android-studio-rabbit-2-run-claude-and-codex-inside-the-ide", "markdown": "https://wpnews.pro/news/android-studio-rabbit-2-run-claude-and-codex-inside-the-ide.md", "text": "https://wpnews.pro/news/android-studio-rabbit-2-run-claude-and-codex-inside-the-ide.txt", "jsonld": "https://wpnews.pro/news/android-studio-rabbit-2-run-claude-and-codex-inside-the-ide.jsonld"}}