Google Open-Sources ARTEMIS for Android Automation, Claims 99%+ Benchmark Result

The tool lets coding assistants operate phones and emulators from plain-English requests. Its two modes make developers choose between quicker execution and more thorough checks.

By 2 min read
Google Open-Sources ARTEMIS for Android Automation, Claims 99%+ Benchmark Result
Google Open-Sources ARTEMIS for Android Automation, Claims 99%+ Benchmark Result

Listen to this story

The audio brief

About 1:40
0:001:40
Read transcript
Google has open-sourced ARTEMIS, a tool that lets AI coding assistants control Android phones and emulators from plain-English instructions. An assistant can build and install an app, move through its screens, and return screenshots, logs, or other diagnostic evidence—bringing device testing directly into a coding workflow. ARTEMIS is available in Google’s public GitHub organization under the Apache 2.0 license. It offers a command-line interface, a web console, a Python software-development kit, and a Model Context Protocol server. That server can connect with tools including Codex, Claude Code, Antigravity, and Windsurf. The important design choice is between speed and verification. Flash is the quicker, reactive mode for routine, predictable interface work. It skips planning, checkpoint checks, final reports, and ADB shell access. Pro is slower, but adds pre-action safety checks, planning, recovery when an action is blocked, diagnostics, and final review. Google says it can handle workflows longer than 100 steps. The tool generally uses visible interface elements, but can fall back to coordinates and visual locating for custom interfaces. Its first task installs an ARTEMIS Accessibility Helper, while UIAutomator2 provides an alternative interface-hierarchy backend. Google reports more than 99 percent completion across AndroidWorld’s 100-plus tasks and 20-plus apps. That is a company benchmark claim, not proof of the same reliability on real devices or custom test suites. The key question now is whether that speed-versus-evidence tradeoff holds up in developers’ own workflows.

Story brief

3 key points

Google’s Pixel-Test-Engineering Fusion team released ARTEMIS under Apache 2.0, giving AI coding assistants an inspectable way to control Android devices, install builds, navigate apps, and collect evidence. Its Flash profile favors fast, predictable actions, while Pro adds planning, safety checks, recovery, verification, and diagnostics for longer workflows. Google reports more than 99% completion on AndroidWorld’s...

  1. 01

    ARTEMIS supports CLI, web console, Python SDK, and MCP integrations with clients including Codex, Claude Code, Antigravity, and Windsurf.

  2. 02

    Pro can handle workflows exceeding 100 steps; Flash omits planning, checkpoint verification, final reports, and ADB shell access.

  3. 03

    The first task installs an ARTEMIS Accessibility Helper; UIAutomator2 provides an alternative interface-hierarchy backend.

Google has open-sourced ARTEMIS, a tool that lets AI assistants carry out plain-English instructions on connected Android phones and emulators. It brings device testing into coding workflows, allowing an assistant to build and install an app, navigate screens, and return screenshots or diagnostic material.

ARTEMIS is published in Google’s public GitHub organization under the Apache 2.0 license. Its built-in Model Context Protocol server connects compatible development environments to a device-control service; the repository names Antigravity, Codex, Claude Code, and Windsurf among supported clients.

The tool runs Android testing workflows and everyday device tasks from natural-language instructions. It uses available interface elements where possible, then can use coordinates and visual locating for custom interfaces. ARTEMIS is available through a command-line interface, web test console, Python SDK, and integrations that collect replays, screenshots, Logcat output, and device diagnostics.

Flash is intended for routine, predictable interface work. Pro is built for cases where the agent must inspect targets before acting, track a plan, and recover from blocked actions; the repository says it can handle workflows of more than 100 steps. The differing time estimates make the tradeoff concrete: more control and evidence come with a slower device loop.

What teams can inspect in the release

  • A first task installs an Artemis Accessibility Helper on the phone; UIAutomator2 is available as an alternative interface-hierarchy backend.
  • The repository claims more than 99% task completion on AndroidWorld, Google Research’s benchmark of more than 100 multi-step tasks across more than 20 apps.

The AndroidWorld result is a company-reported benchmark claim, not proof that every real-world phone workflow will work as reliably. Still, the public code gives developers a way to examine whether ARTEMIS’s split between speed and verification fits their own device tests.

Sources

  1. github.comGitHub - 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. Created by Google's Pixel-Test-Engineering (PTE) Fusion team.

Loading discussion...