Android Studio Coding Agents: Bring Your Own Agent
Google made Bring Your Own Agent stable in Android Studio Rabbit 2 on October 28, 2026. Android Studio coding agents are no longer limited to Gemini. You can pick from over 40 agents, including Claude Agent, OpenAI's Codex, JetBrains' Junie and Google Antigravity. You can also switch between them in the same window. Google previewed the feature in the Rabbit 2 canary on September 24, 2026. Setup takes a sign-in. The catch for anyone switching agents is memory: each agent arrives with only its own.
What is Bring Your Own Agent, in one sentence?
Bring Your Own Agent (BYOA) is an Android Studio feature that runs any Agent Client Protocol (ACP) coding agent in the IDE with its project and device tools.
The Rabbit 2 release notes describe it as connecting your preferred agents "using the open Agent Client Protocol (ACP)." Claude Agent and Codex are two of the three defaults in the agent picker. Google's own Antigravity is the third. The rest come from a list under Settings > Tools > AI > Agents, which Google's Android Dev Summit post calls a marketplace. You can install more than one agent and switch between them as you work.
The protocol's client list also includes Zed and JetBrains IDEs. Our ACP explainer covers how its sessions work.
How does BYOA differ from Gemini's built-in agent?
Gemini Agent Mode stays in Android Studio. BYOA adds outside agents next to it.
| Built-in Gemini agent | A BYOA agent (Claude Agent, Codex and others) | |
|---|---|---|
| Who builds the agent | Google, inside Android Studio | The agent's vendor, connected over ACP |
| Cost | "No extra charge," per the release notes | Your own subscription, enterprise plan or API key, subject to the provider |
| Agents to choose from | One | Over 40, per Google |
| Android tools | Device deployment, screenshots, Logcat and more, per Google's feature list | The same IDE tools, passed in by Android Studio |
| Project instructions | AGENTS.md files and Prompt Library rules, both documented for Gemini | Each agent loads its own project skills, slash commands and memory |
Sources: Google's release notes and preview announcement. The instruction-file row comes from its AGENTS.md guide and its Rules guide.
How do you install a coding agent in Android Studio?
Update to the stable Rabbit 2 release (version 2026.2.2). Then open the agent window and pick one of the three default agents. For anything else, go to Settings > Tools > AI > Agents and browse the list. Each agent asks you to sign in or paste an API key.
Google does not publish the full list of 40-plus agents on its pages. It also does not say which catalog backs the marketplace. The public ACP Registry is one place to see what exists. It lists Claude Agent as an "ACP wrapper for Anthropic's Claude" and Codex as an ACP adapter for OpenAI's assistant.
Is Claude Agent in Android Studio the same as Claude Code?
It is a separate session. Google's agent workflow guide says Android Studio "starts its own instance of the agent through ACP." It does not attach to a Claude Code session already open in your terminal. The Studio session runs on your own Claude sign-in or API key.
Google's preview post says agents in Android Studio "automatically load project skills, slash commands, and memory." So each agent arrives with its own setup. Google documents nothing more specific per agent.
What can a BYOA agent see and do inside Android Studio?
Per the release notes, Android Studio passes the agent the full project graph and its build diagnostics. They also get emulator management and shell execution. The Android SDK tools are on the list too. Google's October posts name a few more, from Logcat to Compose Preview rendering.
Google's pitch is token efficiency. The preview post says the agent gets the project graph over ACP and can then filter to the files it needs. That should cost fewer tokens than an agent exploring a Gradle project from scratch. Google has not published numbers for this.
Rabbit 2 also adds a skill manager that ships with Google's Android and Firebase skills, for jobs like moving XML Views to Jetpack Compose. Skills you add yourself go in a .agents/skills or .android-studio/skills folder, per the skills guide.
What does it cost to use Claude or Codex in Android Studio?
Android Studio charges nothing extra for BYOA. You pay the agent's provider. Google says "both enterprise and consumer AI plans are supported," subject to the agent provider. The release notes also mention personal API keys.
For Gemini, Google points you to the Antigravity agent over the built-in one. Its preview post says Antigravity gets the latest Gemini models and a larger usage quota. You can sign in with a Google AI Pro or Ultra plan, or pay per token with a Gemini API key. Teams on Gemini Enterprise can use either the built-in agent or Antigravity.
What should you read in the fine print?
Every claim about speed and cost savings here comes from Google. We found no independent benchmark of agents running through BYOA versus the same agents in a terminal.
The AGENTS.md and Rules pages describe how Gemini reads them. Gemini scans the current folder and every parent folder for AGENTS.md files. Rules live in the Prompt Library, and project rules are saved under the .idea folder. Google does not say whether an ACP agent receives either one. Claude and Codex read their own instruction files, so a rule saved only in Gemini's Prompt Library may never reach them.
Google's preview post sells switching as a feature: "If one agent runs out of quota," it says, "you can have another agent take over." Google says nothing about carrying over what the first agent learned.
The bottom line
If you build Android apps and already pay for Claude or Codex, try them in Rabbit 2, where they get IDE tools a terminal agent lacks. Put shared project rules in a checked-in AGENTS.md, since a rule saved only in Gemini's Prompt Library may never reach Claude or Codex. Expect each agent to start with only its own memory.
How do you keep project memory when you switch agents in Android Studio?
Picture a Monday in Android Studio. Claude Agent spends the morning on your Gradle upgrade and learns that one module needs a pinned plugin version. At 2 p.m. it runs out of quota, so you switch to Codex, as Google suggests. Codex loads its own memory and has never heard about the pinned plugin.
memU works on the same gap between agents, in a different place. It gives Claude Code and Codex, running in their own apps, one shared memory. It has not been tested with the agents Android Studio launches.
| Agents in Android Studio BYOA | memU | |
|---|---|---|
| Who holds what an agent learned | Each agent loads its own skills, slash commands and memory | One memory backend that every supported host reads. What Claude Code learned, Codex can retrieve |
| How shared rules reach every agent | A checked-in AGENTS.md. Gemini's Prompt Library rules may not reach Claude or Codex | A standing instruction in each host's instruction file, such as CLAUDE.md or AGENTS.md, tells the agent to retrieve before answering |
| Skills | Android and Firebase skills from Google, plus folders you fill yourself | Skills distilled from your own agent history, as readable Markdown |
| Where it runs | Inside Android Studio, which starts its own agent instance | Claude Code and Codex in their own apps, per memU's support matrix. Android Studio is not listed |
Junie and Antigravity are not on memU's list either. For agents without a dedicated adapter, memU's generic memu-agent adapter detects whether it can record that agent's sessions and inject retrieval. Run its detect step before you count on memU for an agent Android Studio launched.
If you use one agent for all your Android work, its own memory plus a checked-in AGENTS.md is enough. memU fits when you split work between two supported hosts. Think of Claude Code one day and Codex the next. More on that split is in our post on sharing memory across coding agents.
To try the hosted version, get an API key from memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is
YOUR_MEMU_API_KEY.
The self-hosted setup is in the memU GitHub repository. More on the design is at memu.pro.
FAQ
Can I use Claude Code in Android Studio?
Yes. Android Studio Rabbit 2 offers Claude Agent, an ACP wrapper for Anthropic's Claude, as one of three default agents. Sign in with your Claude plan or an API key. Android Studio starts its own instance, separate from any Claude Code session in your terminal.
Which Android Studio version supports Bring Your Own Agent?
Android Studio Rabbit 2, stable since October 28, 2026. BYOA first appeared as a preview in the Rabbit 2 canary on September 24, 2026.
How many coding agents does Android Studio support?
Google counts over 40 ACP-compatible agents. Its examples put Junie and Google Antigravity next to Claude Agent and Codex. Any agent that speaks the Agent Client Protocol can connect.
Is Bring Your Own Agent free?
Android Studio charges nothing for it, and the built-in Gemini agent stays at no extra charge. Third-party agents bill through your own subscription or API key with that provider.
Do agents in Android Studio share memory?
Google says each agent loads its own project setup, memory included. It does not describe any shared memory between agents. A checked-in AGENTS.md is the simplest shared layer for written instructions.