moonoops/Droy

Droy is an experimental Android AI agent built on DeepSeek Harness (DSH). It combines a ChatGPT-style chat interface with an "Android Use" toolset — modeled on Computer Use — so the agent can open apps, read page content, take screenshots, tap, swipe, type and verify results. It is aimed at people who want desktop-agent-style task execution on a phone, not those looking for a hosted service: the agent runtime runs on the device after setup.

What a real workflow looks like

You describe a task, and the agent plans it through DSH, which handles planning, model calls and the conversation engine. DSH then calls Android Use tools, backed by native Android executors, to act on the phone. Documented examples include installing and configuring GKD, or setting a scheduled task to open DingTalk for attendance check-in. Steps needing human input — login, verification — are handed back to the user. Caveats are concrete: success depends on app rules, login state, device permissions and background process availability.

Setup paths and their requirements

The recommended path uses a desktop agent that can run local commands (e.g. Codex, Claude Code or DSH) over USB ADB, following the installation guide; that installer supports macOS, Linux and Windows with Python 3.8+ and ADB installation details. The alternative is direct installation of the official signed APK followed by the on-phone setup guide. Either way, post-setup use does not need a connected computer.

Device requirements are ARM64 Android 11 / API 30 or later with an up-to-date system WebView, and the agent is developed on a rooted Redmi K90 Ultra. Without Root, accessibility services provide basic interaction and authorized Shizuku can enable some extra operations; Root enables enhanced capabilities, and personal memory requires Root plus execution-isolation qualification. Other devices and permission combinations need separate compatibility checks. A cloud model must be configured with your own API key; vision models are optional, since text-only models work through page text and control structure, while screenshot understanding requires image input support in the chosen model and API provider capabilities.

Data, risk and licence

The README warns that Droy is experimental and that model errors, defects or system differences may still cause unintended actions, data loss or disclosure. Task context, page text, screenshots and attachments may be sent to the selected model, with the provider's terms and charges applying; voice recognition is local but recognized text goes to the conversation model, and notification analysis can send notification text to the model. Personal memory is off by default, with separate source and processing authorizations. Tasks can be stopped and capabilities changed in settings. The README asks users to disable notification management before uninstalling so changed settings can be restored.

Droy's own code is AGPL-3.0-only, so source-sharing obligations apply; third-party components keep their own licences README.