ARTEMIS

• Asking Claude Code or Codex to build an APK, install it on a connected phone, walk through the login flow, and return screenshots
• Reproducing a reported bug on a real device and collecting Logcat output and a recording
• Running long exploratory or stability sessions with the Pro profile and a written report
• Adding natural-language device checks to a pytest suite through the Python SDK
• Exploring a live app first so a coding agent writes UI test code based on verified interaction paths

ARTEMIS fills a gap most coding agents leave open: they can write and build mobile code but cannot tap through the app on a phone. Exposing a real device through MCP means an assistant can build, install, and exercise a flow, then report back with screenshots and logs. The split between a fast Flash loop and a slower Pro workflow with a separate Checker lets you trade speed for verification per task. Expect some setup work: you need ADB, a device, and model keys, visual coordinate grounding depends on a Gemini Embodied Reasoning model, and the project is only about seven weeks old. Android teams already using AI coding assistants have the most to gain.

About ARTEMIS

ARTEMIS connects to an Android device or emulator over ADB and works through the phone's interface the way a person would: reading the screen, tapping, typing, and moving between apps. It finds elements through accessibility hierarchies and OCR first, then falls back to coordinates and visual models for custom Canvas, Compose, and Flutter interfaces. There are two execution profiles. Flash is a fast observe-and-act loop, typically 3 to 5 seconds per step, for routine tasks, with no task plan, checkpoint verification, or ADB shell. Pro is a multi-agent graph at roughly 15 to 40 seconds per step: a Planner keeps a Markdown plan with checkpoints, an Operator executes it with the full toolset (including ADB diagnostics and video analysis), each individual action passes a pre-execution check against the live UI, and a read-only Checker verifies checkpoints and reviews the result against the original goal. You can use it four ways: a local web console with live screen mirroring and replays, the `artemis run` CLI, a lightweight Python SDK client that calls an ARTEMIS host over HTTP from pytest or CI, or an MCP server with tools such as `mobile_run_task` and `mobile_diagnose`. The setup script (macOS, Linux, or Windows) installs ADB, scrcpy, FFmpeg, and Python dependencies and can register the MCP server and a testing rules file with Antigravity, Cursor, Claude Code, Codex, Windsurf, VS Code, Cline, Roo Code, and OpenClaw. That lets a coding agent build an APK, install it, exercise a flow on a real phone, and return screenshots and Logcat output. You supply model API keys for Google, OpenAI, Anthropic, OpenRouter, or xAI, but the default configuration uses Gemini models, and its comments say the object detector that handles visual coordinate grounding must run on a Gemini Embodied Reasoning model. The project reports a 99%+ completion rate on Google Research's AndroidWorld benchmark. The main limits: it is Android-only for now (iOS is on the roadmap), it installs an accessibility helper app on the device, and it is a young repository created in August 2026 with no tagged releases yet, so expect fast change.