- What it is
- An open-source agent, licensed Apache 2.0, that operates Android phones and emulators from natural-language instructions — built mainly for testing apps.
- How far along
- ResearchPrototypePilotShippingEverywherePublic on GitHub since Aug 13. 121 commits from six contributors, but no tagged release yet.
- Why it matters
- An AI coding assistant can now build an app, install it on a real phone, test it and report back — in one request.
- Still unproven
- Its "99%+" score on Google Research's AndroidWorld benchmark is the project's own claim.
- On your phone
- It installs an accessibility helper. The project says it "listens only on the phone itself and sends nothing elsewhere."
What it does
The project's one-line description says it plainly: "Let AI assistants and test suites use real phones like a human." Give Artemis an instruction in plain language and it operates a connected Android phone or emulator to carry it out — its demo plans driving routes in Google Maps, totals the durations, then opens YouTube to play a song.
It reads the screen three ways: through Android's accessibility layout, with text recognition, and with visual models for custom interfaces built in tools like Flutter or Jetpack Compose. That last part needs a model that can see; the project lists Gemini, Claude, GPT-4o and Qwen-VL.
Google published the repository on August 13, 2026. Some recent coverage presented it as a new release; the commit history shows it has been public for about eleven weeks.
Two speeds
- Flash takes about 3–5 seconds per step: a single model looks at the screen and acts. It is quick, but has no plan, no pre-checks and no final report.
- Pro takes about 15–40 seconds per step and uses three roles. A Planner keeps a written task plan, an Operator carries it out, and a read-only Checker verifies the result. Every action passes a "Safety Net" check against the live screen before it runs. The project says Pro handles workflows of more than 100 steps.
Built to plug into coding assistants
Artemis includes an MCP server — the standard way AI tools connect to outside capabilities — and its setup script offers to configure Antigravity, Cursor, Claude Code, Codex, Windsurf and VS Code automatically. The README's example of what that makes possible:
"Build the latest changes into an APK, install it on the connected device, open the login screen with a test account, verify if there are any unexpected popups after login, and return screenshots of the final page."
There is also a Python client for existing test suites such as pytest, and a web console with live screen mirroring and replays.
What it installs on your phone
The first task on a device installs the "Artemis Accessibility Helper," which shows as a notification and as an entry under Accessibility settings. The project says it "listens only on the phone itself and sends nothing elsewhere," and it can be removed with one command or replaced with Android's standard UIAutomator.
Accessibility services can see everything on the screen — that is how this works. Use a dedicated test device or emulator, not a phone signed into your personal or work accounts.
The numbers
The project reports "99%+ task completion" on AndroidWorld, a Google Research benchmark of more than 100 multi-step tasks across 20+ apps. That figure comes from the project itself; we found no independent reproduction.
As of October 29, the repository had about 15,700 stars and 1,700 forks on GitHub, 121 commits from six contributors, and no formal release — the most recent commit was on September 29. It includes code from Minitap's open-source mobile-use project, also under Apache 2.0.
What's on the roadmap
The project lists four plans: an Android Studio plugin, iOS support, small vision models that run on the device itself, and voice control. These are intentions, not features.
Sources
- google/artemis — GitHub — README, license and commit history, read Oct 29, 2026
- AndroidWorld benchmark — Google Research, GitHub
- Published.