# ToolNeuron

An encrypted, privacy-first on-device AI app for Android that runs chat, vision, retrieval-augmented generation, voice and a sandboxed plugin runtime against local GGUF models, with no Google Play services, telemetry or cloud dependency.

ToolNeuron chats against any compatible GGUF model with streaming output, multi-turn conversation, optional thinking mode, and per-turn tokens-per-second, time-to-first-token and peak-memory metrics; vision support uses colocated `mmproj` projector files, with image bytes passed as file descriptors rather than raw bytes across the app's internal interface. Its RAG pipeline indexes PDF, DOCX, XLSX, PPTX, ODT, EPUB, RTF, Markdown, HTML, JSON, XML, CSV and plain text, with source bytes content-addressed so a document can be reattached to any chat.

Voice runs through sherpa-onnx with streaming text-to-speech and tap-to-toggle speech-to-text, supporting VITS, Piper and Whisper. A separate process runs an HTTP server with OpenAI-shaped endpoints, bearer-token authentication, rate limiting and an audit log, alongside a bundled Material 3 web UI. A plugin store installs sandboxed Android modules with their own Compose UI from a public Hugging Face repository.

An earlier tool-calling and Termux integration was removed in a scope change; models, chats, documents and key material all stay on the device.

Website: https://github.com/Siddhesh2377/ToolNeuron