Trending: On-device modelsSearch
iHeartGeek
iTECH

Android CLI hands your AI agent a real Pixel 10 Pro

Google has brought Android Device Streaming into its command-line tooling, letting an agent spin up a physical phone, deploy a build and pull logs without one sitting on the desk.

The Android skills and Android CLI lockup beside the Android robot mascot on a dark green panel

Google has added Android Device Streaming to Android CLI, its command-line tool for Android development, so that an agent can reach real hardware without anyone owning a rack of test phones. The same release adds a batch of Android skills, including a deep dive on the Wear OS Compose Material 3 skill.

Real handsets, over a terminal

Through the CLI an agent can now connect to a remote physical device over a secure ADB-over-SSL link and treat it as though it were plugged in over USB. Google says that covers spinning devices up, deploying builds, collecting logs and traces, and capturing screenshots headlessly, all from the terminal. The company's own example connects to a Pixel 10 Pro. Developers link a project, then ask the agent to list the remote devices available and pick the one they want.

Skills, and how to install them

Android CLI is built to work with whichever assistant a developer prefers, whether that is Android Studio, Google's Antigravity, or third-party agents such as Claude and Codex, because the skills are deliberately environment-agnostic. Setup is a short set of commands: running android init installs the android-cli skill, android skills list shows the official set, android skills add wear-compose-m3 --project=. drops a single skill into one project root, and android skills update --all refreshes the lot.

The Wear Compose Material 3 skill carries the design decisions that make or break a watch app: round viewports, rotary input, ambient display mode, power consumption, the TransformingLazyColumn list, and the AppScaffold and ScreenScaffold containers. Google points to the engineering team at FotMob, which used it to modernise an existing Wear M3 app, migrating lists to TransformingLazyColumn with ScreenScaffold content padding and adding ListHeader titles, SurfaceTransformation on cards and buttons, theme typography and Wear previews. The changes compiled first time and were checked on an emulator for scrolling, rotary input, edge morphing and right-to-left layouts.

Google is also pointing developers at its longer argument for why skills should be written to be deprecated, rather than accumulating as permanent scaffolding around an app.

Our opinion

Device Streaming in a CLI is a small change with an outsized effect on agent work. The fiddliest part of an autonomous Android loop has never been writing the code, it is the physical handset sitting on a desk, tethered to one machine, with one OS version, and now an agent can rent the exact model it needs for the twenty minutes it needs it.

The more interesting bet is on skills rather than commands. Google is packaging its platform knowledge as separate, versioned modules that any agent can load, which quietly concedes that the model will never ship knowing what a rotary bezel implies for touch targets. That is a sensible division of labour, and it also means the skills can be revised and retired as the platform moves, instead of being baked into weights nobody can patch.