Google has published Artemis, an open-source framework that lets AI coding assistants operate a real Android phone the way a human tester would: tapping, scrolling, reading the screen, and reporting back. The project connects to Claude Code, Codex, Antigravity, Cursor, and Windsurf through a native Model Context Protocol server, so an agent can install a build, walk through a login flow, and pull screenshots and Logcat output without a human touching the device.
The stakes are practical rather than theoretical. Mobile QA has stayed stubbornly manual even as agentic coding tools took over backend and web testing, because phone UIs mix native views, custom Compose and Flutter widgets, and transient popups that break simple scripts. Artemis targets that gap directly, pairing accessibility-tree data with OCR and visual models so it can locate elements in interfaces that do not expose clean labels.
Google’s own benchmark claim is the headline number: Artemis reports a completion rate above 99 percent on AndroidWorld, a Google Research benchmark spanning more than 100 multi-step tasks across over 20 apps. That is a vendor-published result on the vendor’s own benchmark, run by the same organization that built both the tool and the evaluation suite. The repository does not cite an independent third party replicating the score.
Architecture splits the work into two profiles. Flash is a fast, reactive loop, roughly three to five seconds per step, meant for routine deterministic UI tasks with no planning overhead. Pro is slower (fifteen to forty seconds per step) and runs a multi-agent graph: a planner keeps a living task list with checkpoints, an operator executes it with full tool access including ADB diagnostics, and a separate checker verifies each milestone against the original goal before closing out. That split mirrors a pattern showing up across the agent tooling market this year: cheap, high-volume execution for simple actions, paired with a slower, verification-heavy path reserved for anything that needs to be trusted.
On first run, Artemis installs a lightweight accessibility service on the target device to read screen layout, a design choice that avoids hijacking the OS-level UiAutomation connection other testing tools rely on. Google’s documentation states the helper only listens locally and sends nothing off the device, though that claim, like the benchmark score, comes from the project itself rather than an outside audit.
The roadmap listed in the repository includes an Android Studio plugin, iOS support, and on-device lightweight vision models for local execution, none of which has shipped yet. The project is built on source code from Minitap, a mobile-automation startup, and is licensed under Apache 2.0, meaning enterprise teams can fork and modify it without a commercial license fee.
For engineering teams already running Claude Code, Codex, or Antigravity against a codebase, Artemis is a low-friction way to extend that same agent into mobile QA rather than maintaining a separate Appium or Espresso pipeline. Anyone evaluating it should run their own regression suite before trusting the 99 percent figure on a production app, since Google has not published task-level failure data outside the AndroidWorld set.
Google published the Artemis project and its documentation on GitHub; details reflect the repository as of September 15, 2026.
