# Android Studio Opens to Claude, Codex and 40 Other Coding Agents

> Source: <https://explainx.ai/blog/android-studio-rabbit-2-bring-your-own-agent-40-agents-android-cli-agent-apis-dev-summit-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-10-28 00:00:00+00:00

At its Android Developer Summit on October 28, 2026, Google made **Android Studio Rabbit 2** a host for other companies' coding agents. Google's [Bring Your Own Agent post](https://developer.android.com/blog/posts/ai-agents-android-studio-integration?ref=explainx) says the IDE now supports "over 40 coding agents", including Claude Agent, OpenAI's Codex, Junie and Google's own Antigravity. The same day Google turned the Android CLI into an official plugin for Claude Code, Codex and Antigravity and previewed APIs that let apps expose functions to agents.

The move matters because Google owns the Android toolchain. Letting rival agents in, with access to the emulator, build system and Logcat, is a bet that Android's advantage is the platform, not a single assistant. Details below come from Google's posts, relayed by [The Decoder](https://the-decoder.com/google-opens-android-development-to-third-party-ai-agents-and-unveils-new-gemini-apis/?ref=explainx); performance figures are Google's own.

## TL;DR: what Google announced

| Area | What shipped | Status | 
|---|---|---|
| Android Studio Rabbit 2 | Bring Your Own Agent, 40+ agents via the Agent Client Protocol | Stable | 
| Change Review | Analyzes pending AI-generated changes, explains them step by step | Announced | 
| Android CLI | Official agent plugin for Claude Code, Codex, Antigravity | Available | 
| Compose previews in terminal | Agents check UI without an emulator | Available | 
| Android Skills | 30+ documentation packs for agents | Available | 
| Android Bench 2.0 | Leaderboard on multi-day dev projects | Announced | 
| Agentic API (ML Kit GenAI) | On-device agent loop with Gemini Nano 4 | Alpha | 
| AppFunctions API | Apps expose functions to agents through MCP | Beta | 
| Connected Apps Platform | Links apps to Search and the Gemini app | Preview | 

## How does Bring Your Own Agent work?

Rabbit 2 connects to agents through the open Agent Client Protocol, so any compliant agent can plug in. Google's post lists the reasons it gives developers: a preferred agent, the cost efficiency of a particular one, or a provider chosen to match an enterprise agreement. Agents get access to project structure, symbol resolution, the emulator, build systems and Logcat.

To enable one: open the agent window and select an agent, or go to Settings, Tools, AI, Agents; authenticate with that agent; submit your first prompt. Google does not say in the post how usage is billed. You pay whichever provider you authenticate with, so check that vendor's plan and limits.

If you already work with agents in a terminal, this is a different interface to the same engines. Our guide to [loop engineering for coding agents](https://explainx.ai/blog/loop-engineering-coding-agents-claude-code-guide-2026) covers how long-running agent sessions are structured, which matters when an IDE starts embedding them.

## What does the Android CLI add for agents?

The CLI is where Google quantifies its claims:

- **Compose previews in the terminal** let an agent check UI changes without booting an emulator. Google's internal tests reported UI tasks finishing 77% faster and using up to three times fewer tokens.
- **LeakCanary integration** analyzes heap dumps on the computer instead of the device. In a Google A/B test with Claude Opus 5.5, fixing memory leaks was nearly twice as fast with about 25% fewer tokens.
- **Physical device streaming** , from Google Labs, lets agents investigate bugs that only appear on specific hardware.
- **Other hooks** include the Android Performance Analyzer for system traces, Firebase Crashlytics crash reports and the R8 compiler for optimization checks.

These are vendor-run tests with no published methodology in the coverage we read. They are plausible, since giving an agent a cheap way to verify its work reduces retries, but run your own measurement before assuming 77%.

## Android Skills: documentation packs for agents

Google also published more than 30 Android Skills, packs of documentation that tell an agent how to do specific jobs: checking Google Play policy before submission, permission security, adaptive layouts for foldables and the new Googlebook laptops. If the format is unfamiliar, see our [complete guide to agent skills](https://explainx.ai/blog/what-are-agent-skills-complete-guide). The practical effect is that an agent no longer has to guess Play policy from memory, which is where hallucinated advice usually causes rejected builds.

## Agent APIs for apps: AppFunctions, A2UI and Gemini Nano 4

The second half of the announcement is about what your app looks like to an agent:

- **Agentic API (alpha)** in ML Kit GenAI runs on-device with Gemini Nano 4. It executes functions, saves chat history, manages the orchestration loop and includes safety filters.
- **AppFunctions API (beta)** uses the Model Context Protocol as a tool-calling layer so apps can share functions with local or system-wide agents. For background on the protocol, see our[MCP guide](https://explainx.ai/blog/what-is-mcp-model-context-protocol-guide) .
- **Agent-to-UI (A2UI) for Jetpack Compose (alpha)** lets an agent show interactive UI built from your app's own components instead of plain text.
- **Connected Apps Platform (preview)** links apps to Google Search and the Gemini app through Android MCP or cloud-hosted MCP servers. Google says more than 50 partner integrations, including Adobe, Peloton and Experian, are live in Gemini, with web search rollout planned in the coming weeks.

This echoes what Google has been doing on the consumer side, such as [Gemini Personal Intelligence on Android Auto](https://explainx.ai/blog/gemini-personal-intelligence-android-auto-gmail-photos-maps-opt-in-what-it-does-2026) and [opening Home control to Claude and ChatGPT over MCP](https://explainx.ai/blog/google-home-mcp-claude-chatgpt-2026).

## Other announcements in the same keynote

Jetpack Compose 1.13 arrived, and the Jetpack XR SDK, XR emulators and the Engine Hub for Unity, Godot and Unreal are now stable. Google also showed the XREAL AURA, wired Android XR glasses at $1,099, available to order in the US, Canada and South Korea with shipping from mid-November. Those are hardware details outside the AI-coding story, so we only note them.

## Why would Google let rival agents into its own IDE?

Three reasons are visible in the announcement itself. First, developers already pick agents by habit and by company contract: many teams standardize on one provider through an enterprise agreement, and an IDE that only speaks to Gemini loses those teams. Google's own BYOA post names that scenario directly.

Second, Google gets to define the interface. By exposing the emulator, Logcat, project symbols and build tools through a protocol, Google makes Android Studio the place where agents learn Android, and Android Skills and the CLI plugin extend that into the terminal. Whichever model wins, the tooling underneath is Google's.

Third, it feeds Google's measurement story. Android Bench 2.0 scores models on multi-day development projects, and a stable way to run any agent in the IDE makes it easier to compare them under the same harness. Whether Google's own models win its own leaderboard is exactly what outside observers will check; we saw similar skepticism around internal coding claims in our coverage of [Bloomberg's report on Google's internal Gemini coding skepticism](https://explainx.ai/blog/google-gemini-4-bloomberg-internal-coding-skepticism-2026).

## A realistic first session

Here is how a cautious team might evaluate this over a week, based only on features Google documents:

1. **Day one:** install Rabbit 2, open the agent window and connect the agent you already pay for. Ask it to explain an unfamiliar module, which is low risk and shows how well it reads your project symbols.
2. **Day two:** give it a small bug with a failing test. Watch whether it uses the emulator and Logcat itself or asks you to paste output.
3. **Day three:** switch agents on the same task. This is the point of BYOA, and the only honest way to compare cost and quality on your code.
4. **Day four:** run the Android CLI in a CI job with a read-only task, such as a lint or leak check, and review the output before enabling anything that edits code.
5. **Day five:** compare token use and review time. If the 77% and 25% savings Google reports do not show up on your project, that is useful data too.

## What this means for Android developers

1. **Try your agent in the IDE** , but keep a terminal session with the Android CLI for CI. The same plugin serves both.
2. **Use Change Review on agent output.** Google's tool is meant to explain and flag AI-generated changes before you accept them.
3. **Watch cost, not just speed.** Android Bench 2.0 reports functionality, time and cost, which is the right trio; wait for third-party results.
4. **Think about AppFunctions early.** If agents become a way users reach your app, the functions you expose are your new public interface. Treat them like an API with auth and abuse limits. Agents with real device access can do damage if they are tricked;[AgentBeam](https://agentbeam.com/?ref=explainx) is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

## What is not yet clear

Google's post also does not expand the protocol name or describe how permissions are scoped when an agent touches the emulator or a connected physical device, so security-minded teams should ask before enabling device streaming broadly.

- Pricing and quotas for third-party agents inside Studio depend on each provider.
- Android Bench 2.0 results for specific models were not in the coverage we read.
- Alpha and beta APIs may change before general release.

## Related reading

*Details are accurate as of October 28, 2026; alpha and beta features may change.*