OpenGUI: an Android agent that taps through real apps

OpenGUI puts an AI agent behind the touchscreen. Instead of calling app APIs, it reads what is on the phone, decides on a sequence of taps and swipes, and executes them through Android's accessibility layer — the same channel screen readers use.

Planner and operator are different models

The backend runs a supervisor graph that breaks a task down and an executor subgraph that handles the individual gestures. Each half can point at a different provider, and that split is where the project makes its cost argument.

For the visual work the repo recommends Doubao VLM first and Qwen VLM as the practical second choice. OpenAI vision models are listed as capable but more expensive, Grok vision as experimental. Pairing Qwen 3.6 Plus for text roles with Doubao Pro for vision, the project claims roughly 10x to 15x lower cost than an all-Opus setup. That figure is the publisher's own; no independent measurement backs it.

Jobs can arrive from a chat window

A standby gateway accepts tasks over Feishu, Telegram, Discord or a REST endpoint, so a run does not have to start at the desk. The design target is long workflows — jobs that run for hours, track their own progress and resume after a break.

Setup is not casual. You need Android 11 (API 30) or newer, USB debugging on, AccessibilityService granted, overlay permission and a battery-optimization exemption. ADB, the project states, is used only to install and launch the APK and to set up local port forwarding; it does not root the device or touch the Android system.

The license clock runs to 2030

OpenGUI ships under the Business Source License 1.1. The code is readable and modifiable, but production deployment requires a commercial license until the Change Date of April 29, 2030, when it converts to Apache 2.0. Teams evaluating it should treat it as source-available tooling, not as open source they can build a product on today.

What nobody has verified yet

The repo is candid about its gaps: execution quality depends on the model, the app's own interface, network conditions and how long the task runs, and the long-run reliability that the design promises still needs testing outside the lab. Ready-to-run examples and benchmarks are thin.

Popularity is a snapshot rather than a trend — about 1,700 stars, 69 forks, 23 open issues and 192 commits at the time of reading. Only the project's own page was available as a source for this report, so every performance claim above is the publisher's, not a third party's.