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.
PocketStrike is a privacy-first mobile AI workstation. It combines zero-cloud native on-device GGUF inference (ARM NEON accelerated) │ with frontier cloud LLMs, multi-step ReAct agent tools, Model Context Protocol (MCP) servers, self-evolving cross-session memory, and FLUX neural image │ generation.
⚡ 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.
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
Generate images directly in any conversation using natural language:
# 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
git clone https://github.com/your-username/PocketStrike.git
cd PocketStrike
flutter pub get
2. Verify Code & Run Tests
# Static analysis (0 warnings / errors)
flutter analyze
# Unit & widget test suite
flutter test
3. Run in Debug Mode on Device
flutter run -d <device_id>
4. Build Final Production Release APK
flutter build apk --release
The optimized release APK will be generated at:
build/app/outputs/flutter-apk/app-release.apk
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.
P
PocketStrike
AbuZar-Ansarii·146 estrellas·
PocketStrike is a privacy-first mobile AI workstation. It combines zero-cloud native on-device GGUF inference (ARM NEON accelerated) │ with frontier cloud LLMs, multi-step ReAct agent tools, Model C
PocketStrike is a privacy-first mobile AI workstation. It combines zero-cloud native on-device GGUF inference (ARM NEON accelerated) │ with frontier cloud LLMs, multi-step ReAct agent tools, Model Context Protocol (MCP) servers, self-evolving cross-session memory, and FLUX neural image │ generation.
Herramientas
Herramientas
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Preguntas
¿Cómo instalo PocketStrike?
Copia en esta página la configuración stdio para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.