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

	Minimum	Recommended
Android	10 (API 29)	12+
RAM	6 GB	8–12 GB
Storage	4 GB free	10 GB free
CPU	ARM64 or x86_64	Snapdragon 8 Gen 1+

---

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

Use case	Model	Size
Quick test	Qwen3.5 0.8B Q4_K_M	~600 MB
General use	Qwen3.5 4B Q4_K_M	~2.8 GB
Power users	Qwen3.5 9B Q4_K_M	~5.5 GB

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

Tool	Status
Upscaling	Ready
Segmentation (MobileSAM)	Ready
Depth estimation	Model pending
Style transfer	Model pending
LaMa inpainting	Model pending

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:

Plugin	What it does
Web Search	Search the web and scrape content
File Manager	List, read, create files
Calculator	Math expressions and unit conversion
Notepad	Save and retrieve notes
Date & Time	Current time, timezone conversion, date math
System Info	RAM, battery, storage, device details
Dev Utils	Hash, encode, format, text transforms

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

Layer	Technology
Language	Kotlin, C++ (JNI)
UI	Jetpack Compose
Text inference	llama.cpp
Image inference	LocalDream (SD 1.5)
TTS	Supertonic (ONNX Runtime)
Database	Room + UMS (custom binary format)
Encryption	AES-256-GCM, Android KeyStore
DI	Dagger Hilt
Async	Kotlin Coroutines + Flow

Modules

Module	Purpose
app	Main Android application
ums	Unified Memory System — binary record storage with JNI
neuron-packet	Encrypted RAG packet format with access control
memory-vault	Legacy encrypted storage (read-only, used for migration)
system_encryptor	Native encryption primitives
file_ops	Native file operations

---

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