PocketStrike

# 📱 PocketStrike > **Executive Multi-Provider & 100% Offline AI Workstation for Android & Mobile** > *Bridge the gap between Cloud Frontier LLMs, Local On-Device GGUF Inference, Unlimited AI Image Synthesis, Model Context Protocol (MCP), and Self-Evolving Cross-Session Memory.* [](https://flutter.dev) [](https://dart.dev) [](https://github.com/ggerganov/llama.cpp) [](https://modelcontextprotocol.io) [](LICENSE) --- ## 📁 Repository Structure ``` PocketStrike/ ├── android/ # Native Android configurations, ProGuard rules, & NDK bridges ├── lib/ │ ├── app/ # Theme, Glass Tokens, & Router configuration │ ├── core/ # Drift SQLite Database, DAOs, Models, Notifications & Services │ │ ├── db/ # Drift tables & queries │ │ ├── models/ # Provider-agnostic chat models & metrics │ │ └── services/ # Android notification & background download service │ ├── shared/ # Glassmorphic UI components, code blocks & custom icons │ └── features/ │ ├── agent/ # ReAct Agent Engine, Steps, & Built-in Tools │ ├── chat/ # Chat UI shell, input bar, message cards & speed badges │ ├── conversations/ # History drawer & thread management │ ├── local_models/ # Native GGUF inference runtime, model downloader & presets │ ├── mcp/ # MCP protocol client (Streamable HTTP & SSE) & marketplace │ ├── providers/ # Multi-provider gateways (OpenAI, Claude, Gemini, DeepSeek, Ollama) │ ├── settings/ # Local models, MCP, Storage, API keys, & Hermes memory UI │ └── telegram/ # Telegram Bot autonomous background relay service └── test/ # Unit and widget test suite ``` --- ## 🌟 Key Features & Highlights - ⚡ **100% Offline On-Device Native GGUF Inference**: Run state-of-the-art quantized SLMs locally on your phone using ARM NEON multi-threaded CPU acceleration. Zero cloud dependency, zero data leaving your device. - 🔥 **Dedicated Uncensored & Standard SLM Suite**: Curated 1-tap presets including **SmolLM2 135M/360M Heretic**, **Llama 3.2 1B Abliterated**, **Qwen 2.5 0.5B**, **DeepSeek R1 Distill 1.5B**, and **Llama 3.2 3B**. - ⚡ **Real-Time Token Generation Speed (`⚡ X.X tok/s`)**: Live generation throughput metric displayed dynamically in the bottom-right corner of the AI response container. - 🎨 **Unlimited AI Image Generation Engine**: Create infinite images from natural language (`/image <prompt>`) with custom aspect ratios (`1:1`, `16:9`, `9:16`, `4:3`, `3:4`), interactive zoom/pan viewer, and 1-tap gallery downloads. - 📥 **In-App Model Downloading with Live Notification Progress**: Download model weights straight from HuggingFace to your phone's storage with background Android notifications and 1-tap **"Load into RAM"**. - 📐 **Adaptive Auto-Fitting Chat Layout**: User message bubbles automatically shrink-wrap to their content length (e.g., single-word prompts fit tightly without stretching). - 🤖 **Autonomous ReAct Agent Engine**: Multi-step tool execution with step-by-step reasoning logs, real-time status chips, and configurable confirmation safety gates. - 🔌 **Model Context Protocol (MCP) Marketplace**: Native Streamable HTTP & SSE transport support with 11+ built-in servers (Brave Search, GitHub, Puppeteer, PostgreSQL, Memory Graph, Hardware Specs, etc.). - 🌌 **Hermes Self-Evolving Memory System**: Persistent `USER.md`, `MEMORY.md`, and `SOUL.md` files that autonomously learn your personality, preferences, and context across sessions. - 📁 **Direct Mobile Storage Management**: File explorer and document tools to read, write, search, and manage files in `/storage/emulated/0/Download` and app documents. - 🤖 **Telegram Autonomous Remote Relay**: Turn your phone into a 24/7 background AI assistant that receives prompts and replies over Telegram. - 🎨 **Cyberpunk Glassmorphism UI**: Isolated backdrop blur, dark/light themes, syntax-highlighted code blocks, and haptic feedback. --- ## 🏗 Architecture Overview ```mermaid graph TD A[Flutter Glassmorphism UI] --> B[Riverpod State Controllers] B --> C[Chat & Agent Coordinator] C --> D[Local AI Engine] C --> E[Cloud Multi-Provider Gateway] C --> F[ReAct Agent Engine] C --> G[Hermes Memory Core] C --> H[Neural Image Synthesizer] D --> D1[Native llama.cpp ARM NEON Bridge] D --> D2[Offline GGUF Weight Store] D --> D3[Background Download Manager] E --> E1[OpenAI / Claude / Gemini / DeepSeek / Ollama / OpenRouter] F --> F1[Built-In Mobile Tools] F --> F2[MCP Server Transport HTTP / SSE] G --> G1[USER.md / MEMORY.md / SOUL.md] H --> H1[Turbo Diffusion Engine & Local Caching] ``` --- ## 🧠 Curated Local Model Suite | Model Name | Parameters | Quantization | RAM Footprint | Category | Key Capability | | :--- | :---: | :---: | :---: | :---: | :--- | | **SmolLM2-135M-Instruct-heretic** | 135M | `Q4_K_M` | ~150 MB | 🔥 Uncensored | Ultra-compact, near-zero RAM footprint, zero alignment refusals | | **SmolLM2-360M-Instruct-Heretic** | 360M | `Q4_K_M` | ~350 MB | 🔥 Uncensored | Balanced reasoning & instruction-following without guardrails | | **Llama-3.2-1B-Abliterated** | 1.2B | `Q4_K_M` | ~950 MB | 🔥 Uncensored | Unrestricted creative writing, coding, and general assistance | | **Qwen2.5-0.5B-Instruct** | 0.5B | `Q4_K_M` | ~450 MB | 💬 Standard | Multilingual instruction following & fast summaries | | **Llama-3.2-1B-Instruct** | 1.2B | `Q4_K_M` | ~950 MB | 💬 Standard | Official Meta 1B model, high quality formatting & reasoning | | **DeepSeek-R1-Distill-Qwen-1.5B** | 1.5B | `Q4_K_M` | ~1.3 GB | 💬 Standard | Distilled reasoning model with chain-of-thought output | | **Llama-3.2-3B-Instruct** | 3.2B | `Q4_K_M` | ~2.1 GB | 💬 Standard | Comprehensive complex multi-step reasoning & coding | | **SD-Turbo-Image-Gen** | Diffusion | `FP16` | ~1.8 GB | 🎨 Image Gen | Fast single-step local image generation | --- ## 🎨 Image Generation Commands Generate images directly in any conversation using natural language: ```bash # Square (1:1) /image cyberpunk neon samurai meditating in rain # Widescreen (16:9) /image futuristic electric sports car speeding through tokyo 16:9 # Mobile Wallpaper (9:16) /image breathtaking bioluminescent alien forest at night 9:16 # Natural Language Phrases draw a cute fluffy orange cat wearing an astronaut helmet generate an image of a medieval library filled with glowing crystals ``` --- ## 🚀 Getting Started & Build Instructions ### Prerequisites - **Flutter SDK**: `>= 3.29.0` - **Dart SDK**: `>= 3.12.0` - **Android SDK & NDK**: Installed with compileSdk 37, minSdk 26 - **Physical Android Device** (connected via USB/ADB) or Emulator ### 1. Clone & Setup Dependencies ```bash git clone https://github.com/your-username/PocketStrike.git cd PocketStrike flutter pub get ``` ### 2. Verify Code & Run Tests ```bash # Static analysis (0 warnings / errors) flutter analyze # Unit & widget test suite flutter test ``` ### 3. Run in Debug Mode on Device ```bash flutter run -d <device_id> ``` ### 4. Build Final Production Release APK ```bash flutter build apk --release ``` The optimized release APK will be generated at: `build/app/outputs/flutter-apk/app-release.apk` --- ## 🔒 Security & Sandboxing - **On-Device Storage Control**: Configurable root directories in **Settings > Storage & Permissions**. - **Human-in-the-Loop Safeguards**: Irreversible operations (`delete_file`, `overwrite_file`) require explicit confirmation depending on your chosen policy (`Destructive Only`, `Always Ask`, or `Autonomous`). - **Hardware AES-256 Storage**: All cloud API keys and bot tokens are encrypted in hardware-backed platform storage via `flutter_secure_storage`. - **R8 / ProGuard Safety**: Native JNI symbol preservation guarantees native model execution in release builds without runtime reflection crashes. --- ## 📄 License PocketStrike is open-source software licensed under the **[MIT License](LICENSE)**.