A project by one person, filed under Google's name: ARTEMIS gained 16.7k stars in 85 days, riding the Google name, one striking number and a passing wave

ARTEMIS is an open-source agent that drives Android phones with plain-language instructions and installs into Claude Code, Codex and Cursor in one click. It went public in early September and reached 16,682 stars in 85 days. But it was almost entirely written by one person, part of its code came from an open-source project by a small French company, and that project has only 3,400 stars. This piece covers how its stars came about, what you can learn from it and what you can't, and whether this area is still worth building in.

"We wrote this, how did it end up under Google's name"

In 2026 Sep 11, the Minitap team opened google/artemis in their office. Their CEO wrote a blog post that day titled "I expected better from Google", recording his first reaction: "What the fuck? We wrote this."

He listed specific evidence: the adb tunnel code for Android devices was identical to their mobile-use; a sub-agent called Hopper had word-for-word identical prompts (the name came from an engineer who likes Minecraft, tossed out on a whim); and the WhatsApp example task, sending New Year's greetings to Alice, Bob and Charlie, was the same, down to a bug from an older version. An earlier pyproject.toml listed three Minitap engineers by name, and a force push in August swapped those names for another author.

The next day, a Hacker News post titled "Google stole open source code…" reached 180 points. That same day (9-12), ARTEMIS gained 1,191 stars, its biggest single-day gain (per Trendshift's daily data). The negative coverage didn't cool the buzz; it pushed it further. Later the repo added Apache 2.0 attribution under third_party/mobile_use.

What it is

A Python Android automation framework, Apache-2.0. Say "plan two routes in Google Maps, total the travel time, then open YouTube and play a Coldplay song", and it taps through each step on a real device or emulator. The README's first screen shows exactly this demo GIF.

It has two modes. Flash is a fast loop that acts as it sees the screen; the README puts each step at 3–5 seconds. Pro has three agents, planning, executing and checking, cooperate, checking before each action, and ends with a test report. Elements are located through the accessibility tree and OCR, and it switches to a vision model for self-drawn UIs like Compose and Flutter. It doesn't bring its own model; Gemini, Claude, GPT and Qwen all plug in.

Setup is a single ./start.sh: it installs adb, scrcpy, ffmpeg and the Python dependencies, then asks whether to write the MCP config and a mobile-testing rules.md into 8 IDEs, including Antigravity, Cursor, Claude Code, Codex and Windsurf. One badge in the README reads "AndroidWorld 99%+ SOTA", and AndroidWorld is Google Research's own benchmark. The 99% figure is self-reported by the project; I found no independent reproduction.

How well it's growing

Sources: Trendshift daily table, GitHub daily records. Over the last 7 days +1,040, the first 7 days +1,862, and it's slowing. Code hasn't changed since Oct 1; contributor somew1nd (GitHub profile says Google) made 109 commits, everyone else single digits, and only 61 PRs were merged in 3 days.

Where the stars came from

Wave one was "Google just dropped this." No official blog, no launch event. Even the HN launch post got only 2 and 4 points. What actually lit the fuse was Jaewoong Eum's (@github_skydoves, Android GDE) tweet in the early hours of Sep 10: "Google published an AI Android app automation tool." Then came GitHub Trending: 9/10–9/13 overall #6, #5, #3, #9, and #1 on the Python chart. On TikTok, "GitHub recommendation" accounts followed: github.signals had a 42k-view video, howtowebdev 36k. In the Chinese-speaking world, @Gorden_Sun posted a tweet with 21k views, a cnblogs post titled "Googles's new AI automation open-source tool is so strong that Appium is panicking" (夯爆了,谷歌又一强大AI自动化开源神器,Appium慌了) ran, and the repo itself ships a README_CN.

Wave two was global AI content creators. From mid-to-late October, short videos in Portuguese, Thai, Indonesian, Italian, Hebrew and Arabic rolled out in turn, each getting 3–40k views. Portuguese creator andrereis0027 10-19 had one at 34k, captioned "11 mil estrelas"; Thai creator tonsr_ 10-25 had one at 43k. Many were lead-gen accounts with "comment ARTEMIS and I'll DM you the link." At the same time, a dozen or more English AI news sites (10-09 through 10-27) wrote about it at once. There's no sign of a single source. It looks like creators across markets each spotted "Google gives away free AI that operates your phone for you" as a good topic and posted it one after another (inference).

Three things had to come together:

• The brand. The repo sits under the google org, which has 2.16M stars in total. Every repost headline read "Google open-sourced..." The same code under mobile-use has only 3,367 stars.
• A shareable number. "99%" appears in the badge and the description, backed by Google's own benchmark. Nearly every Chinese and English creator headline used this figure.
• A wave. In the second half of 2026, everyone was wiring MCP into Claude Code and Codex. "Let coding agents reproduce bugs and run tests on a real phone" was an existing need for Android developers. The README names Antigravity, Codex and Claude Code together, so users of each felt it was for them.

Functionally it does a bit more than mobile-use or mobile-mcp (multi-agent verification, one-click IDE install, Logcat and screen recording, a pytest SDK), but most people reposting it never installed it. The stars came from the story, not the features.

Who it's competing with

This category was already crowded. ARTEMIS wasn't the first, and it didn't have the most features:

Under skydoves' tweet, someone replied "Another day, another agent for mobile devices." A TesterArmy person also pushed their own e2e framework in the comments. The commercial real-device cloud Kobiton published an article riding the buzz.

If it were you

The verdict: most of ARTEMIS's stars came from the words "google/", which you can't get. What you can learn is the packaging, but the "mobile coding agent" spot already has five or six projects with tens of thousands of stars. There's something to learn, but copying it isn't worth it.

Who it suits: engineers at big companies or well-known orgs who can publish a project under the org name, or people who already have a working mobile agent and want to learn how to package and launch it.

Who it doesn't suit: solo developers who want to build an "AI operates Android" framework from scratch and expect to hit ten thousand stars. Without a big-company name, the same code would land at roughly mobile-use's level.

How to do it:

• Copy the packaging, not the positioning. The README's first screen has three pieces: a slogan in the style of "Let AI assistants use real phones like a human," a benchmark badge with a cited source, and a GIF that anyone can follow (map → play a song on YouTube).
• The install script also writes the MCP config and agent rules into popular IDEs, so people who repost it can run it within 30 seconds. Each IDE's community also gets a new topic.
• If you do build in this space, go after what it does poorly: real iOS devices (phone-harness just started on this), fast replay-cached tests (what TesterArmy pointed out in the comments), or an independently reproduced benchmark. Small and precise, not another all-in-one framework.
• Credit third-party code clearly from day one. ARTEMIS got plenty of stars this time, but its reputation took a hit. Without a big-company name, a single blog post saying "they copied us" is enough to kill your project.

How to get attention: follow the same path, but make the first hop someone you can actually reach: find developers whose posts carry weight on X in your field (for Android, the GDE crowd), and have them post first. Time it to the two or three days of GitHub Trending. Prepare one image that says "free + does the work for you" and a GIF, then let multilingual short-video accounts translate and repost it on their own.

When to move: the wave of MCP plugins that wire peripherals into coding agents is still going, but the phone slice is already packed within a few months. To claim a niche (iOS, or one specific testing scenario), you have to move before Google or Apple puts official tools into Android Studio or Xcode. No one knows when that will happen, so earlier is better.

Signs you've misread it: if Google keeps ARTEMIS maintained as a real product and ships it into Android Studio, the spot is gone for good. Conversely, if it sees no updates after November and its issues go unanswered, and Droidrun and mobile-mcp also fail to capture real developer usage, then "one well-built small tool" still has room.