wangzhishou / OneBox

# OneBox (万宝盒)
 **English** | 简体中文
 A free AI agent & toolbox for Android: say what you need and the built-in agent drives
 90+ in-app tools to get it done — plus image & document tools, productivity and daily utilities.
## Screenshots
## Download
- **International**: Google Play · Official site
- **中国**: 万宝盒官网 · 小米 / 应用宝 / OPPO / vivo / 华为应用商店搜索「万宝盒」
## AI Assistant: More Than Just Feature Count
Plenty of toolbox apps are long on features. OneBox's built-in AI assistant saves you from hunting for tools and memorizing steps — just say what you want:
> "Turn these images into a PDF."
> "Log this bill screenshot into my ledger."
> "Make me a to-do list."
> "Look up today's stock prices online."
It understands, acts, and reports back — in one flow.
- **Bounded by its own permissions**: everything the agent does stays inside the app's own Android sandbox and declared permission set — it never touches contacts, SMS or call logs; local tools run entirely on-device, and only the conversation with your chosen model goes online.
- **An agent that drives the tools**: the AI agent can invoke 90+ in-app local tools (PDF processing, image editing, file management, bookkeeping…), chaining multi-step tasks across tools — executed locally with per-tool timeouts and iteration limits.
- **Picture-in-picture robot, no more waiting around**: long agent task? Just leave the app — with an active AI session it drops into system picture-in-picture, where the mini-window robot keeps you posted on progress and one tap brings you back to the result. Native PiP, zero permissions required.
- **One chat, many models**: talk to multiple mainstream LLMs from a single entry, with custom AI engines and prompt management.
- **Bring your own key — use it for free**: drop in any OpenAI-compatible API key (including free tiers like Gemini or OpenRouter's free models) and the full assistant + agent is yours at zero cost. Custom engines, endpoints and tokens are fully open in the Google Play build and in builds from this repo.
- **Models that run on the phone — fully offline**: download a local model (Gemma 3 1B or Qwen3 0.6B) inside the app, or import your own `.litertlm` file, and chat without network, account or token cost — inference runs entirely on-device via LiteRT-LM (Google Play and FOSS builds only).
- **Transparent by source**: every line is public — how data is processed locally and what never leaves the device is something you can verify, not a promise you have to trust.
## Features
OneBox ships **104 feature modules**. Local tools run entirely on-device; only the conversation with your chosen AI model goes online. The tables below list every module — names refer to the `feature/*` directories in this repo.
### AI & Assistant
| Module | Feature | Description |
| --- | --- | --- |
| `feature/ai` | AI Chat & Agent | Multi-model chat, dual-model duels, prompt execution, an agent that invokes in-app tools, and offline on-device models via LiteRT-LM (Google Play / FOSS builds),