Google released a new Android CLI on April 16, a companion Skills repository, and something called the Android Knowledge Base, all aimed at AI coding agents building native Android apps outside of Android Studio. The toolkit works with Gemini CLI and Google's own Antigravity, sure. But it also works with Anthropic's Claude Code and OpenAI's Codex, which is the more interesting detail.
What Google actually shipped
Three things, really. The Android CLI wraps what used to be scattered across sdkmanager, avdmanager, adb, and parts of gradle into a single terminal interface with commands like android create, android emulator, and android run. The Skills repo is a collection of markdown files, one per workflow, that agents pull in automatically when a prompt's metadata matches the skill. And the Knowledge Base, triggered by android docs, lets agents fetch current documentation from developer.android.com, Firebase, and Kotlin docs at query time rather than relying on whatever the model memorized last year.
Google's announcement post, co-authored by group product manager Adarsh Fernando and senior staff software engineer Esteban de la Canal, frames the suite as a way to eliminate guesswork when agents work outside Android Studio. Anyone who has watched a coding agent try to install an SDK component through forty lines of sdkmanager output understands why this exists.
The initial Skills release covers Navigation 3 setup and migration, edge-to-edge support, AGP 9 and XML-to-Compose migrations, and R8 config analysis. Not the sexiest list. Also not a coincidence. These are precisely the areas where LLMs trained a year ago will confidently generate code that doesn't compile against current Android.
About those numbers
Google claims the CLI cuts LLM token usage by more than 70% during project and environment setup, and gets those tasks done roughly three times faster than an agent stumbling through the legacy Android SDK toolchain. Those are the headline figures. They come, per the blog post, from "internal experiments."
Internal experiments. Designed by the team that built the CLI. Against a baseline of agents fumbling through sdkmanager and adb directly. I'd be surprised if the numbers came back any other way.
That doesn't mean they're wrong, exactly. The underlying claim is plausible: a consolidated CLI that emits predictable, structured output is easier for a language model to drive than six separate tools printing progress bars and ANSI color codes. But 70% and 3x are specific numbers, and Google hasn't published the methodology, the agent tested, the prompts, or the task set. Until someone outside Mountain View runs this against Claude Code or Codex on a fresh laptop, treat the figures as marketing until verified.
The strategic part
What's more telling than the performance claims is the list of supported agents. Google is actively shipping a toolkit that makes Claude Code and Codex better at building Android apps. Their stated goal is that high-quality Android development should be possible everywhere, and they specifically mean on tools that aren't Google's.
Read that cynically and it's a concession. Developers are already using Claude Code and Codex in large numbers; Google can fight that or feed it. Google, apparently, is feeding it, on the theory that owning the tooling layer matters more than losing some model-layer traffic to competitors. The blog post ends with a pitch to eventually move projects into Android Studio for "premium" finishing work, which is the part where Gemini comes back into the picture.
Granted, there's also a genuine platform argument. Android's addressable market only grows if Android is easy to target, and right now it's measurably harder to build a decent Android app with a generic coding agent than an iOS one. Lowering that friction helps the platform whether the developer pays Google for anything or not.
An Anthropic standard, on a Google tool
Small detail worth flagging: Android Skills uses the open Agent Skills format, which is maintained by Anthropic. The SKILL.md convention, the metadata-based triggering, the progressive disclosure pattern, all of it originated with Claude. Google's repo essentially ships a standard Anthropic published months ago. That's the kind of cross-vendor cooperation that wasn't on anyone's 2024 bingo card.
The install script defaults to ~/.gemini/antigravity/skills if no agent directory exists. It also respects .claude/skills/. The tooling assumes you might be running any of them, or several at once.
What's missing
A few things. There's no iOS analog, of course. Cross-platform developers on Flutter or React Native gain little from this release directly, though presumably the Skills pattern could be extended by the community.
Also absent: any independent verification of the performance claims, a target date for moving from preview to stable, and any acknowledgment of what happens when the Knowledge Base's latest guidance contradicts an app's existing codebase. The Skills demo in the announcement shows an agent cleanly migrating XML layouts to Compose. Real migrations are messier than that.
Android CLI is available in preview starting April 16. The Skills repo is public on GitHub. There's no stated date for a stable release.
