wangzhishou/OneBox
AI EngineeringAndroid AI agent and toolbox that runs 90+ in-app tools from natural language and opens 100+ file formats.
What It Is
OneBox is an Android app that combines a built-in AI agent with an oversized toolbox. Instead of hunting through menus, you describe a task ("turn these images into a PDF", "log this bill", "make a to-do list") and the agent selects and chains 90+ on-device tools to complete it. It also acts as a viewer for 124 file extensions, covering images, PDF, Office documents, Markdown, code, audio, video and ZIP archives.
The project is a fork of ImageToolbox; the image-processing core stays close to upstream under Apache-2.0, with additional AI, document and utility modules. It targets Android users who want one app for file viewing, image editing, conversion, OCR and light productivity, with optional AI assistance.
Quick Start
Install from Google Play or the official site; a Chinese site also exists at wanbaohe.com. On first launch, open Settings → AI engines to add an OpenAI-compatible API key or endpoint, or download a local model (Gemma 3 1B or Qwen3 0.6B, roughly 585–615 MB) for offline chat. You can then ask the assistant to perform a task, or open a file from the system file manager and let OneBox route it to the right viewer or editor.
To build from source, use JDK 17 and the bundled Gradle Wrapper 9.8.2, shallow-clone the repo, then run:
git clone --depth=1 https://github.com/wangzhishou/OneBox.git
./gradlew :app:assembleGoogleArm64Debug
The README warns that builds from the repository do not contact production servers and hosted content stays empty until you configure your own backend. See run.md for additional build commands.
On-Device AI and Limits
Chat, translation and photo text recognition can run locally through LiteRT-LM and ONNX Runtime. Chat uses a .litertlm model (Gemma 3 1B or Qwen3 0.6B); translation uses Tencent Hy-MT2 1.8B in 20 languages; OCR uses a 26 MB PP-OCRv6 pack with SHA-256 verification. These paths need no account or network, but the local chat and translation runtimes are absent from domestic-channel builds, while OCR is included in every build.
Long documents, tables and image understanding still require cloud models. The agent is bounded by the app's declared Android permissions and never accesses contacts, SMS or call logs; only the conversation with your chosen model goes online. Release builds need a real keystore.properties for signing; debug builds work without one, though third-party service features remain disabled.
HotRepo's take
OneBox has 1,191 stars and 204 forks, and presents a very large surface: 107 feature modules, 124 file extensions and a long module table in the README. The repository does include source code, build variants and a documented fork history, so this is not a thin wrapper. The main practical question is scope: a one-person project promising an AI agent, image editor, file viewer and dozens of utilities will have uneven polish across modules, and the domestic/overseas build split changes which AI features are available. The README's contributor section itself asks for device testing and translation help, which is consistent with that breadth-versus-depth tradeoff. There is no usable HotRepo ranking or star-curve history to assess recent momentum, so the observation rests on the repository contents.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- Verlintas/NovaBAICLocal-first Android AI agent that calls 58 device tools to read the screen, automate UI, manage files and run scheduled tasks.
- guzhou079-arch/deepseek-harness-androidAn Android APK that bundles DeepSeek Harness with screen automation, virtual display, scheduled tasks and 27 skills, without Termux or root.
- tapilot-ai/tapilotAn experimental Android app that executes plain-language phone tasks using a vision model and accessibility control.
- AbuZar-Ansarii/NeedleA local Termux agent that turns plain-English requests into Android phone controls and offline queries.
- droidrun/mobilerunAn LLM-agnostic framework that lets AI agents inspect, tap, swipe, and type on real Android and iOS devices via CLI or Python.