ARTEMIS is Google's open-source tool that lets AI assistants and test frameworks control real Android phones or emulators through natural language commands. Instead of writing traditional test scripts, you tell it what to do—"Open Settings, find Battery and tell me current level"—and it executes the task by observing the screen, identifying UI elements, and performing actions. The system hits 99%+ completion on Google's AndroidWorld benchmark, which tests 100+ multi-step tasks. It works through three main interfaces: a web console with live screen mirroring, integration into coding IDEs via Model Context Protocol (MCP), and a Python SDK for CI/CD pipelines. Setup is straightforward—clone the repo, run a one-click script that installs dependencies, connect an Android device with USB debugging enabled, and you're testing. The tool uses a two-speed approach: "Flash" mode runs quick reactive tests in 3–5 seconds per step, while "Pro" mode does deeper reasoning for complex scenarios. It finds UI elements by checking element indices first, then falls back to visual detection or coordinates if needed. The MCP integration means you can plug real phones directly into Antigravity, Claude Code, Windsurf, and other AI IDEs—the AI assistant becomes a mobile test engineer that explores the interface before acting, catches blocking errors, and recovers gracefully. Google includes a rules file that teaches AI assistants how to test like a senior engineer, covering things like timing compensation and when to use visual-first versus coordinate-based locating. Getting started requires minimal setup. The one-click launcher auto-detects and installs ADB, scrcpy, FFmpeg, and Python dependencies, then prompts you to configure MCP for your IDE. You can run tests from the web UI, the command line, or embed them in Python code with the SDK. The web console shows real-time device projection, action trajectories, reasoning logs, and lets you replay executions. For developers, this means testing Android apps without writing Espresso or UIAutomator tests—you describe what you want checked, and the AI handles the interaction.