Opposable

getopposable.comDR 10 Accessible 300 views

Development & Programming Opposable is an Android app that lets AI models control other apps by reading the screen and tapping or typing.

Visit now Favorite

Editor's Review

For reference only

Opposable stands out by running a 2B GUI model entirely on the phone, with no server needed, while still allowing remote models to drive the same loop. The blog details concrete benchmarks, including 7.3 s per step on a laptop CPU and a 236 ms vision encoder on a Galaxy S25 NPU. It suits developers

Site introduction

Opposable is an Android app built around computer use on the phone itself. It reads the screen through the accessibility service and taps and types in any other app, letting AI models operate Android apps directly. The project comes from Reval Labs and is documented through blog posts, recorded runs, and pricing pages.

The core feature is a dual-mode agent loop. Claude, ChatGPT, or Codex can drive the app from a computer, and when none of them is connected, a 2-billion-parameter screen model runs the same loop on the phone with no server involved. The on-device model is GUI-Owl 1.5 2B in 4 bits on llama.cpp.

The blog reports concrete performance numbers. On a Mac M2 CPU with 4 threads, one step takes 7.3 s, down from 10.6 s, at the same accuracy on a 36-screen test set. Half of that step is the vision encoder, which runs in 236 ms on a Galaxy S25 NPU versus 3.9 s on the M2.

The site is aimed at developers, researchers, and technical users interested in on-device AI agents and mobile automation. It is also relevant to anyone evaluating the trade-offs between local and remote model execution for GUI control tasks.

For best use, start with the TL;DR and recorded runs to understand the loop, then check the limitations section before trying it. The blog is a practical reference for on-device agent performance rather than a general consumer guide.