Minitap Accuses Google's Artemis of Copying Its Agent Code

A small Android-automation startup says Google copied its open-source agent code line for line, then rewrote the author list to erase them from it. Minitap, the two-person-plus team behind the Apache-licensed mobile-use project, published a post on September 11 laying out code comparisons between its repo and Google's Artemis, a mobile-automation agent from the Pixel Test Engineering team. The connector code matches almost line for line, the prompt files match word for word, and even a demo script that texts "Happy New Year" to fictional users named Alice, Bob and Charlie shows up in both places. Google has not responded publicly as of publication.

• Minitap says Google's Artemis agent shares near-identical connector code, agent prompts and example scripts with its own mobile-use, an Apache 2.0 project first built in early 2026.
• An earlier Artemis commit listed three Minitap engineers as authors; a force-push in August 2026 replaced that file and removed their names entirely, leaving only a different author credited.
• Minitap says it submitted benchmark results as high as 100% on Google's own AndroidWorld leaderboard between December 2025 and January 2026 and never got a reply; a comparison chart Google later published included two rival tools but left mobile-use out.
• The dispute is playing out entirely in public: a GitHub issue, a Notion timeline with commit hashes, and a Hacker News thread that had cleared 80+ points within hours.

What is Minitap actually accusing Google of?

Minitap built mobile-use as an open agent framework that lets an AI model operate a real Android or iOS device the way a person would: tap, type, read the screen, adapt to what comes back. It runs on Apache 2.0, a license that lets anyone use or fork the code commercially, provided they preserve copyright notices and flag what they changed. Google's Artemis does the same job, built by the Pixel Test Engineering "Fusion" team and pitched as a way to drive a phone straight from Antigravity, Codex, or Claude Code over MCP. Minitap's post lines up three specific files: the ADB tunnel connector that talks to the physical device, the instruction file for an internal sub-agent Minitap calls "Hopper," and a demo script. In each case the founder says the structure, variable choices, and even inherited bugs line up too closely to be coincidence, and points to exact commit hashes on both repos so anyone can check the diff directly.

RelatedFlutter 3.44 adds agentic hot reload and GenUI

Why does the author removal matter more than the code match alone?

Code convergence happens; two teams solving the same automation problem can land on similar shapes. What turns this from "maybe a coincidence" into a documented grievance is the paper trail Minitap says it can point to. An earlier commit to Artemis's pyproject.toml listed three Minitap engineers, Pierre-Louis Favreau, Jean-Pierre Lo and Nicolas Dehandschoewercker, as authors. A subsequent force-push swapped that file for one crediting someone else entirely, with GitHub's own history marking the earlier commit as detached from the branch rather than deleted outright. A force-push does not erase git's record, which is exactly how Minitap says it was able to reconstruct the before-and-after. Whether that credit was ever meant to stay, or was an internal placeholder later cleaned up, is the specific question Google has not answered yet.

What happened with the benchmark and the comparison chart?

Minitap says it had also been trying to get onto Google's own AndroidWorld leaderboard, the benchmark Artemis cites for its 99%+ completion claim. The startup submitted a 91.4% result that was accepted in December 2025, then followed up in January 2026 with runs scoring 94.8% and eventually 100%, task traces included, and says two follow-up emails went unanswered. As of publication the public leaderboard still lists mobile-use at the older 91.4% figure while Artemis shows 99.1%. A comparison chart Google circulated alongside Artemis's launch included DroidRun and a MadeAgents tool also called MobileUse, but not Minitap's project, despite the code overlap Minitap says it can demonstrate.

Who gets hurt if this holds up?

Most directly, a small team that built something a much bigger company's benchmark chart quietly stepped around. Minitap is a startup selling autonomous QA agents for mobile testing, and the mobile-use repo is both its research showcase and, under Apache 2.0's terms, its recruiting pitch: use it, but say where it came from. If a well-resourced competitor can absorb the code, ship a near-identical product with more distribution, and leave the origin off both the credits and the comparison chart, the practical incentive for small teams to open-source their best work gets weaker. More broadly, this lands in the middle of an existing nerves-frayed debate about AI-assisted coding blurring the line between "inspired by" and "copied from," the same worry Ladybird's maintainers cited in June when they closed public pull requests over unreviewable AI-generated contributions.

RelatedGoogle Play Threatens to Delist AnkiDroid Over Donations

What would settle this, and what happens next?

Three things would move this from allegation to resolved fact: Google explaining the pyproject.toml swap, Google adding or crediting mobile-use on the AndroidWorld comparison chart, and an independent read of the actual diffs Minitap linked. Apache 2.0 does not require permission to reuse code, only that attribution and change notices survive the reuse, so the legal bar here turns on whether Google preserved what the license requires, not on whether copying happened at all. Minitap has filed a public GitHub issue on the Artemis repo and published a timestamped Notion record of the commit hashes rather than going the legal-threat route first, a choice that keeps this resolvable with a changelog entry rather than a lawsuit, if Google chooses that path.

Our take

The strongest thing Minitap has going for it here is not the code match, it is the paper trail. Anyone can claim a rival "copied" them; fewer teams show up with exact commit hashes on both sides and a documented force-push that happens to delete their names right as a well-funded competitor's version goes public. That specificity is what separates this from typical open-source griping. It does not prove intent on Google's part, and there are boring explanations for a commit history that looks damning out of context. But the omission from the comparison chart, on top of the unanswered benchmark emails, on top of the authorship swap, is a pattern that deserves a real answer rather than silence. Small teams building the open tooling that larger companies then ship polished versions of is not new. What is newer is how easy git history makes it to document exactly when that happened, down to the commit.

Original analysis by GenZTech. Story developing; Google has not issued a public response as of publication.