ToolNeuron

Offline AI assistant for Android. Run LLMs, generate images, search documents — all on-device. No cloud. No subscriptions. No data leaves your phone.

Download APK · Discord · Report Issue

What It Does

• Text generation — Load any GGUF model (Llama, Mistral, Gemma, Phi, Qwen, etc.) and chat with it locally
• Image generation — Stable Diffusion 1.5 on-device, with inpainting support
• Image tools — Upscale and segment images locally (depth, style transfer, inpainting coming soon)
• RAG — Inject PDFs, Word docs, Excel, EPUB into conversations with semantic search
• Plugins — Web search, file manager, calculator, notepad, date/time, system info, dev utils — all callable by the LLM
• AI memory — The AI remembers facts about you across conversations, with deduplication and a forgetting curve
• Text-to-speech — 10 voices, 5 languages, on-device synthesis
• Encrypted storage — AES-256-GCM with hardware-backed keys for all chat data
• System backup — Export everything as an encrypted .tnbackup file

Requirements

Getting Started

1. Install

Google Play or GitHub Releases.

2. Get a model

From the in-app Model Store (recommended):

1. Open the drawer menu → Model Store
2. Add a HuggingFace repository (e.g. bartowski/Phi-3.5-mini-instruct-GGUF)
3. Pick a quantization and download

Or manually:

1. Download a .gguf file from HuggingFace
2. Use the model picker in ToolNeuron to load it

3. Chat

Select your model, wait for it to load, start typing. Responses stream in real-time.

Recommended models for getting started

Pick Q4_K_M for a good balance between quality and size. Use Q6_K if your device has the RAM for it.

Features

Text Generation

• Any GGUF model works — load via file picker (no storage permissions needed, uses SAF)
• Configurable parameters: temperature, top-k, top-p, min-p, repeat penalty, context length
• Function calling with grammar-constrained JSON output
• Thinking mode for models that support it
• Per-model configs saved to database

Image Generation

• Stable Diffusion 1.5 (censored and uncensored variants)
• Text-to-image and inpainting
• Configurable steps, CFG scale, seed, negative prompts, schedulers

Image Tools

RAG (Document Intelligence)

Create knowledge bases from:

• Files — PDF, Word (.doc/.docx), Excel (.xls/.xlsx), EPUB, TXT
• Text — Paste any text content
• Chat history — Convert past conversations into searchable knowledge
• Neuron Packets — Import encrypted .neuron RAG files

The RAG pipeline uses hybrid retrieval: FTS4 BM25 + vector search + Reciprocal Rank Fusion + Maximal Marginal Relevance. Results are injected into the conversation context automatically.

Encrypted RAGs support admin passwords and read-only user access.

Plugin System

7 built-in plugins the LLM can call during conversations:

AI Memory

Inspired by Mem0. After conversations, the LLM extracts facts about you and stores them for future context. Deduplication via Jaccard similarity, with a forgetting curve so stale memories decay. You can view, edit, and delete memories from the Memory screen.

Text-to-Speech

On-device TTS via Supertonic (ONNX Runtime). 10 voices (5 female, 5 male), 5 languages (EN, KR, ES, PT, FR). Adjustable speed and quality. Auto-speak option reads responses aloud.

Hardware Tuning

Auto-detects CPU topology (P-cores, E-cores) and recommends thread count, context size, and cache settings. Three modes: Performance, Balanced, Power Saver.

System Backup

Export everything to an encrypted .tnbackup file (PBKDF2 + AES-256-GCM):

• Chat history, AI memories, personas, knowledge graphs
• Model configs and app settings
• RAG files and AI models (optional, can be large)

Privacy

• Zero data collection. No telemetry, no analytics, no crash reporting.
• Everything stays on-device. Conversations, generated images, documents, TTS audio — none of it leaves your phone.
• Encrypted storage. AES-256-GCM with Android KeyStore. On supported devices, keys live in the Trusted Execution Environment.
• No storage permissions. Models load through Android's file picker (SAF). The app can't access arbitrary files.
• Open source. Read the code yourself.

Building from Source

Prerequisites

• Android Studio Meerkat (2025.1.1)+
• JDK 17
• Android SDK 36+, NDK 26.x

Build

git clone https://github.com/Siddhesh2377/ToolNeuron.git
cd ToolNeuron

# Debug
./gradlew assembleDebug
./gradlew installDebug

# Release
./gradlew assembleRelease

APKs land in app/build/outputs/apk/.

If you hit NDK issues, make sure NDK 26.x is installed via SDK Manager. For memory issues during build, bump the Gradle heap in gradle.properties:

org.gradle.jvmargs=-Xmx4096m

Architecture

Modules

Contributing

See CONTRIBUTORS.md for the project ecosystem and related repos.

How to contribute

1. Fork the repo
2. Create a feature branch: git checkout -b feature/your-feature
3. Make focused commits with clear messages
4. Test on a real device — emulators don't reflect real performance
5. Open a PR with a description of what you changed and how you tested it

Priority areas

• Bug fixes and stability
• Device compatibility testing (especially mid-range phones)
• Performance improvements
• Documentation and translations
• New plugins

What not to do

• Don't submit untested code
• Don't add cloud dependencies or telemetry
• Don't break offline functionality
• Don't add broad storage permissions

Security

If you find a security vulnerability:

1. Do not open a public GitHub issue
2. Email siddheshsonar2377@gmail.com
3. Include reproduction steps
4. Allow reasonable time for a fix before disclosure

Acknowledgments

• llama.cpp — LLM inference engine
• LocalDream — Stable Diffusion on Android
• ONNX Runtime — TTS inference
• Mem0 — AI memory architecture inspiration
• SillyTavern — Character card format reference
• Apache POI, PDFBox-Android, EpubLib — Document parsing
• Jetpack Compose, Room, Hilt, OkHttp, Coil 3, Jsoup

License

Apache License 2.0 — use it, modify it, distribute it. Attribution appreciated.

Built by Siddhesh Sonar

Star the repo · Report a bug · Join Discord