Google ships LLM dev CLI

Google just shipped a new command-line tool for Android development, and the real story is not “yet another CLI.” It’s that Google is starting to package Android itself for AI coding agents. The new Android CLI is meant to be the terminal-first interface those agents can reliably use for setup, project creation, emulator management, and running apps. That matters because a lot of LLM-powered coding breaks down on boring workflow steps, not on writing snippets of Kotlin. ### What did Google actually launch? On April 16, Google introduced Android CLI, plus “Android skills” and an Android Knowledge Base for agentic workflows. The company framed it very directly: whether you’re using Gemini CLI, Gemini in Android Studio, Claude Code, Codex, or another agent, Android development should work cleanly outside the IDE too. Version 0.7 is the first public release. Agents do better when the environment is predictable and machine-readable. Traditional Android setup has been scattered across separate tools and manual steps — install SDK pieces, create a project, spin up a device, deploy, troubleshoot. A human can muddle through that. An LLM often burns tokens reading docs, guessing commands, and retrying. Google’s pitch is basically: give the model one clean interface instead of a maze. ### What can it do? The core commands are pretty practical. `android sdk install` pulls only the SDK components you need. `android create` scaffolds a new project from official templates. `android emulator` handles virtual devices. `android run` deploys the app. There’s also `android update` so the tool itself can keep up with new capabilities. Google also says the CLI can help agents navigate UI flows on a device, which is a bigger deal than it sounds because testing loops are where a lot of agent workflows stall. ### What’s the headline number? Google says internal experiments showed more than 70% lower LLM token use for project and environment setup, and tasks completed 3x faster than when agents tried to use the standard Android toolset on their own. That does not mean every Android task is suddenly 3x faster in production. It means the scaffolding phase — the annoying setup layer — got much more efficient. But that’s still meaningful, because setup friction is exactly where “vibe coding” often turns into dead ends. ### Is this only for Google’s own models? No — and that’s one of the more interesting parts. Google explicitly named Gemini CLI, Claude Code, and Codex as compatible starting points. So this is less about locking developers into one assistant and more about making Android a first-class target for any coding agent that can use terminal tools. In other words, Google wants the Android platform to be easier for agents to operate, even if the agent came from a rival. ### How does this fit Google’s bigger AI tooling push? It lines up with a broader move inside Google toward CLIs and packaged “skills” for agent development. Just days later, Google Cloud introduced Agents CLI for building and deploying agents on its cloud stack, with the same basic idea — reduce context overload, inject the right capabilities, and let coding agents scaffold projects without inhaling huge amounts of documentation first. Android CLI is the mobile-dev version of that pattern. ### What’s the catch? A faster scaffold is not the same thing as a production-ready app. The CLI can standardize setup and templates, but it does not remove the need for human review, testing on real devices, security checks, or safety guardrails around any LLM-powered feature. Basically, Google is solving the “how do I get started fast?” problem. The “how do I ship something trustworthy?” problem is still yours. Google is turning Android development into something AI agents can operate more natively. That sounds small, but it’s a real platform shift. The less time models spend fumbling with environment setup, the more useful they become for actual building.