Android Studio Rabbit 2 Canary adds Bring Your Own Agent, or BYOA, through the Agent Client Protocol. It lets Android developers run supported external coding agents inside the IDE while preserving Android-specific project, build and emulator context.
The feature is a preview, not a promise that every agent behaves identically. Google names Claude Agent, OpenAI Codex and Google Antigravity among the supported options, plus other ACP-compatible agents. Availability, authentication and billing still come from the agent provider.
What Android Studio BYOA changes
A terminal agent can already edit an Android repository, but it does not automatically understand everything the IDE knows. Android Studio’s integration can provide the agent with the project graph, build diagnostics, Android SDK context, terminal access and emulator controls. The practical benefit is fewer copy-and-paste loops when a change fails because of a Gradle task, manifest issue or device state.
ACP is the connection between the IDE client and the coding agent. It is not MCP: MCP connects a model or agent to tools and data, while Agent Client Protocol carries coding-agent sessions between an editor and an agent backend. It is also unrelated to the older Agent Communication Protocol discussed in our agent protocol comparison.
Requirements
- Android Studio Rabbit 2 Canary 2 or later in the current preview channel
- A supported agent installed or configured on your machine
- A valid provider subscription or API credential where the selected agent requires one
- A project you can build locally before handing changes to an agent
Canary builds are pre-release software. Use a disposable branch or worktree, keep production signing keys out of the agent environment, and review generated changes before merging.
Set up an external agent
- Install the current Android Studio Rabbit 2 Canary build.
- Open Settings > Tools > AI > Agents.
- Select an available agent from the registry or configure an ACP-compatible agent.
- Complete that provider’s authentication flow.
- Open an Android project and start with a bounded task, such as explaining one build failure.
- Inspect the diff, run the relevant Gradle tests and verify the result on an emulator or device.
Do not give an agent broad repository and shell access merely because the IDE exposes those capabilities. Start read-only where possible, restrict secrets through your environment-variable and secret workflow, and require approval before destructive commands.
Claude Agent vs Codex vs Antigravity
| Agent | Choose it when | Check before rollout |
|---|---|---|
| Claude Agent | You already use Anthropic’s coding workflow and want it in Android Studio | Provider plan, context behavior and tool permissions |
| OpenAI Codex | Your team uses Codex for repository tasks and reviewable changes | Authentication, sandbox policy and organization controls |
| Google Antigravity | You want Google’s agent workflow alongside Android tooling | Preview status, model/runtime choice and local versus managed execution |
| Other ACP agent | You need a specialized or self-hosted agent | ACP compatibility, maintenance quality and security model |
The decision should follow your existing governance and code-review process, not a benchmark screenshot. Test the same real task with each finalist and compare build success, unnecessary changes, latency and reviewer effort.
A safe evaluation workflow
Use one representative feature and one representative failure:
- Ask the agent to explain the relevant module and propose a plan without editing.
- Allow a small implementation on a dedicated branch.
- Run unit, lint and instrumentation checks.
- Inspect whether the agent used Android-specific build or emulator context correctly.
- Record corrections a human reviewer had to make.
This turns BYOA into a controlled AI coding-agent workflow, rather than granting a new tool blanket trust. For broader quality gates, connect the result to AI testing and evaluation and your existing pull-request approvals.
Current limitations
- Rabbit 2 is in the Canary channel, so UI and compatibility can change.
- ACP compatibility does not guarantee identical feature support across agents.
- Provider subscriptions, rate limits and data policies still apply.
- IDE context improves Android awareness but does not guarantee a correct build or safe change.
- Enterprise teams still need explicit rules for source access, secrets, telemetry and generated-code review.
FAQ
What does BYOA mean in Android Studio?
Bring Your Own Agent lets you connect a supported external coding agent to Android Studio through Agent Client Protocol instead of being limited to one built-in assistant.
Does Android Studio support Claude and Codex?
Google lists Claude Agent and OpenAI Codex among the agents supported in the Rabbit 2 Canary preview. This does not include a provider subscription, and exact capabilities can vary.
Is Agent Client Protocol the same as MCP?
No. Agent Client Protocol connects an editor client to a coding agent. MCP standardizes access to tools and data. A coding agent can use both for different layers of its workflow.
Is BYOA ready for production teams?
It is useful for evaluation, but the Android Studio implementation is still a Canary preview. Test it on isolated branches, preserve human review and avoid exposing production secrets until your team has validated its controls.