Google lets third-party AI agents into Android Studio: Claude, Codex and 40+ more

Google has stopped reserving Android Studio's AI assistants for its own tools. At the Android Dev Summit, the company introduced Android Studio Rabbit 2 with a Bring Your Own Agent feature. Developers can connect any of more than 40 external AI agents to the code editor. These are programs built on language models that read a project, edit code and run checks on their own. Google also sped up such agents in the terminal, released interfaces for AI features on the smartphone itself, and showed the XREAL AURA smart glasses.

Your own agent in someone else's IDE

Android Studio used to run mostly Google's built-in tools. Now agents connect through the open Agent Client Protocol (ACP). It works as a shared "socket": the editor and the agent negotiate through it, so nobody has to write a separate integration for each one.

Through the built-in marketplace, developers can install and run the following at the same time:

• Claude Agent from Anthropic; • Codex from OpenAI; • Google Antigravity, Google's own agent.

Agents get direct access to the IDE's tools: the project structure, code symbol definition search, the emulator and Logcat (the app's event log on the device).

AI-generated code has a known problem: the edits come out scattered, and a human has trouble making sense of them. Google added Change Review to address this. The tool goes through pending edits, compiles an error report and explains step by step what changed. Version 2.0 of the Android Bench ranking has also been released. It measures how well language models handle multi-day projects and accounts for functionality, time spent and cost.

The Canary test build now includes a migration assistant. It automatically converts old XML layouts to Jetpack Compose, Android's modern toolkit for describing UI in code. Google is also testing Lightbuild, a build system based on a YAML schema. Its files are easy for both humans and agents to read and edit.

How agents can save tokens

Tokens are the unit used to measure the volume of text a model processes. The more tokens, the more expensive and slower the work. That is why Google expanded Android CLI, a command-line utility for people who run agents in the terminal or in automated build pipelines. It is now available as an official plugin for Claude Code, Codex and Antigravity.

New commands let an agent render a Compose preview right in the terminal and check how a UI edit looks without loading an emulator. According to Google's internal tests, agents finished UI tasks 77% faster and used up to three times fewer tokens. The utility can also stream physical test devices from Google's labs, so an agent can debug errors that appear only on a specific phone model.

To find memory leaks, Google built the open-source LeakCanary library into the CLI and into the Android Studio profiler. It analyzes memory dumps on the computer rather than on the phone, so the app does not slow down during the check. In an A/B test with Anthropic's Claude Opus 5.5 model, leaks were found and fixed almost twice as fast, with about 25% fewer tokens.

Agents can also capture system traces through Android Performance Analyzer, pull crash reports from Firebase Crashlytics and check code optimization by the R8 compiler.

Models often don't know the newest APIs because they were trained on older data. Google's answer is a package of 30+ Android Skills, ready-made documentation sets for agents. They help check Google Play rules before publication, work out permissions and adapt layouts to foldable devices and Googlebooks laptops.

Apps as tools for agents, and glasses on top

For running AI locally on a phone, Google is adding an alpha version of the Agentic API to ML Kit, based on the Gemini Nano 4 model. The API calls functions on its own, stores chat history, runs the orchestration loop (where the agent decides step by step what to call next) and provides safety filters.

So that apps can share their functions with agents on the device or across the whole system, Google is adopting the Model Context Protocol (MCP) and launching a beta of the AppFunctions API. An alpha Agent-to-UI (A2UI) interface for Jetpack Compose is built on top of them. Instead of a wall of text, the agent shows an interactive interface assembled from the app's own components.

To get app services into Google Search and the Gemini app, Google is launching a preview of the Connected Apps Platform. Developers connect apps through Android MCP or cloud MCP servers and manage them from a single console. More than 50 partner integrations, including Adobe, Peloton and Experian, already work in Gemini, and they will reach web search in the coming weeks.

On the hardware side, Jetpack Compose 1.13 and the AppState model make it easier to adapt apps to large screens. XR development tools (Jetpack XR SDK, XR emulators in Android Studio, Engine Hub for Unity, Godot and Unreal) have reached stable versions.

Together with the manufacturer XREAL, Google showed the XREAL AURA glasses. According to Shahram Izadi, head of the XR division, they are the first wired smart glasses on Android XR. They weigh less than 95 grams and have transparent lenses with adjustable dimming and a 70-degree field of view.

The compute unit and battery sit in a separate puck-shaped controller, which doubles as a trackpad and runs on the new Qualcomm Snapdragon Reality Elite chip. Users can place up to five Android apps in their field of view or mirror a laptop screen over USB-C. Gemini sees both the real surroundings and the virtual windows: it arranges them, gives tips in games and highlights objects in the room.

The glasses can be ordered in the US, Canada and South Korea for $1099, with shipping starting in mid-November.

For engineers, the takeaway is simple: Android development is no longer tied to a single AI assistant, and you can pick an agent for the task. Token savings in the terminal and ready-made documentation sets lower the cost of agent work. And MCP and the Connected Apps Platform give apps a chance to show up where users are already talking to AI.