// the find
google/artemis
ARTEMIS turns natural-language instructions into reliable Android automation. It automates end-to-end workflows, captures logs, and integrates seamlessly with AI coding assistants such as Antigravity, Codex, and Claude Code. It also achieves 99%+ success rate on AndroidWorld Benchmark.
ARTEMIS is Google's Android automation agent that turns natural-language instructions into real device interactions, exposed as an MCP server (so Claude Code, Antigravity, Codex can drive a phone directly), a CLI, a Python SDK, and a web console. It's aimed at QA engineers and anyone wiring up agentic mobile testing or bug repro into an AI coding workflow, not casual scripters.
The two-profile split is the smart part: Flash is a cheap reactive observe-act loop for routine taps, Pro is a full planner/operator/checker graph with a pre-execution safety net for anything that needs verification — you pick the cost/rigor tradeoff per task instead of one-size-fits-all. Element location combines accessibility hierarchy, OCR, and visual models with a dynamic-first/coordinate-fallback order, which actually covers custom Canvas/Compose/Flutter UIs that accessibility-only tools miss. MCP being a first-class server (not bolted on) means it's genuinely usable as a tool inside an existing coding agent session rather than a separate harness you have to shell out to.
The architecture has sprawled fast: 15+ specialized agents (planner, operator, checker, explorer, diagnoser, log_analyzer, video_analyzer, object_detector, history_analyzer, outputter, summarizer...) plus a separate FastAPI/Angular admin console. That's a lot of moving parts to reason about when a run fails partway through. It installs a persistent accessibility service on the target device on first use — the README's 'sends nothing elsewhere' claim isn't something you can verify without reading the helper's source yourself. The headline 99%+ AndroidWorld number is self-reported on Google's own 100-task benchmark with no independent reproduction, so it's a weaker signal on messy real-world apps than the badge implies. And despite supporting four different vision-capable models per step, there's no token/cost accounting anywhere in the docs — for a Pro-profile run that can go 100+ steps, that's an unknown you eat before you find out it was expensive.