PRIVACY68 Voice Assistant

PRIVACY68 is a 100% private Windows desktop AI assistant featuring on-demand Click-to-Speak, a modern minimalist Control Center, universal multi-provider AI brain orchestration, and agentic OS automation. It pairs zero-key speech recognition (Google Web Speech) with an Agent Brain powered by universal AI APIs (Google Gemini, Groq LPU, OpenRouter, Anthropic Claude, OpenAI, or local Ollama / LM Studio) to execute complex desktop tasks safely through generated, self-correcting Python scripts.

Latest Windows Release: Download PRIVACY68 Installer v1.6.2 from GitHub Releases
Universal LLM Guide: Read Complete Guide to Connecting Any Cloud or Local LLM

---

What's New in v1.6.2

• Universal API & Custom Endpoint Support: Connect any cloud AI key (Gemini, Groq, OpenRouter, Claude, OpenAI) or any custom OpenAI-compatible endpoint (LM Studio, LocalAI, vLLM, DeepSeek, Mistral).
• Intra-Provider Model Cascading: Automatically cascades to sibling models within the same provider on 404/429 errors (e.g. Groq decommissioned model fallback to openai/gpt-oss-120b).
• Autonomous Configuration Healing: Automatically updates runtime .env settings upon failover success to eliminate retry latency on future commands.
• OpenRouter Free Tier (100+ Free Models): Direct first-class support for free open-weights models (openrouter/google/gemini-2.0-flash-exp:free, openrouter/meta-llama/llama-3.3-70b-instruct:free).
• Local Ollama Model Chips: Control Center auto-detects installed local models and displays interactive one-click activation chips.
• 87 Unit Tests Passing: Fully validated test suite (100% pass rate in 13.9s) covering auto-failover, self-healing, security guard, and skill crystallization.

---

Command Routing Architecture

The assistant uses deterministic shortcut dispatching alongside an LLM-powered Agent Brain:

Voice/Text Command
  -> Command Executor (actions/executor.py)
       ├── Pending Security Confirmation ("yes", "confirm", "cancel")
       ├── Built-in Greeting Shortcuts ("hi", "hello", "good morning")
       ├── Direct App Launch Shortcuts (Chrome, Brave, WhatsApp) -> Native Plugins
       ├── GitHub Integration Plugin -> GitHub REST API & local Git operations
       ├── Skill Crystallizer -> 0ms fast-path execution of cached automation
       └── Agent Brain (actions/agent_brain.py)
            ├── Multi-Provider AI Routing (Gemini / Groq / Claude / OpenAI / Ollama)
            ├── Automatic Failover on Rate Limits (Groq 429 -> Gemini Fallback)
            ├── Dynamic Windows Path Resolution (OneDrive/Desktop/Downloads)
            ├── Isolated Code Execution with 30-Second Watchdog
            ├── Self-Healing Error Correction Loop (up to 2 retry attempts)
            └── Formatted Results & UI Telemetry to Control Center & Floating HUD

---

System Requirements & Prerequisites

1. Hardware Specifications

Component	Minimum Specification	Recommended Specification
Operating System	Windows 10 (64-bit, build 19041+)	Windows 11 (64-bit)
Processor (CPU)	Intel Core i3 / AMD Ryzen 3 (AVX2 support)	Intel Core i5/i7 / AMD Ryzen 5/7
System Memory (RAM)	4 GB	8 GB or 16 GB
GPU (Optional)	None (runs on CPU via int8)	NVIDIA GeForce GTX 1650 / RTX 3050+ (6 GB+ VRAM)
Audio Hardware	Standard USB/3.5mm microphone & speakers	Noise-cancelling array microphone
Storage	1.5 GB free disk space	5 GB+ (for local LLMs if stored locally)

---

2. Multi-Provider AI Brain Options

PRIVACY68 offers complete freedom in selecting your reasoning provider:

1. Free Cloud AI (Zero Setup & Fast):
   • Google Gemini: Supports gemini/gemini-3.5-flash with generous free tier keys from Google AI Studio.
   • Groq Cloud: Ultra-fast inference running groq/openai/gpt-oss-120b or qwen3.8-27b at ~200ms latency.
   • OpenRouter: Access 100+ free open-weights models (openrouter/meta-llama/...:free, gemini-2.0-flash-exp:free).
2. Premium Cloud AI:
   • Anthropic Claude: Claude 3.5 Sonnet / Haiku for complex multi-step reasoning.
   • OpenAI: GPT-4o / GPT-4o-mini for robust desktop automation.
   • Custom Endpoints: DeepSeek, Mistral, Together AI via standard OpenAI-compatible base URLs.
3. 100% Offline Local Models (Ollama, LM Studio, LocalAI):
   • Run completely offline with zero internet and zero cloud token usage:
   • Ollama: Connects to http://localhost:11434 or remote Ollama server.
   • LM Studio / LocalAI / vLLM: Connects to http://localhost:1234/v1 or custom base URL.
   • Quick one-click model selection right from the Control Center.
   • For full setup instructions, see the Universal LLM & API Integration Guide.

---

3. CUDA & GPU Acceleration (Optional, for High-Speed Audio Inference)

PRIVACY68 uses Google Web Speech (100% free, zero keys required) for high-accuracy speech transcription with zero background listening. For local offline Whisper transcription, GPU acceleration is supported:

Compute Mode	Audio Inference Latency	Memory Footprint	Setup Required
CPU (cpu / int8)	~350ms – 600ms	~1.2 GB System RAM	Zero setup. Works out of the box on any Windows PC.
GPU (cuda / float16)	~120ms – 160ms	~1.4 GB NVIDIA VRAM	NVIDIA GPU + cuBLAS 12 & cuDNN 9 runtime DLLs.

Zero-Config One-Click Solution:

If an NVIDIA GPU is detected on your PC, the PRIVACY68 Control Center provides a "Download CUDA (~450MB)" button under the Hardware card. Clicking it downloads the exact pre-compiled cuBLAS 12 and cuDNN 9 runtime DLLs directly into:

%APPDATA%\PRIVACY68\cuda\bin\

Upon restart, PRIVACY68 automatically registers these DLLs and switches the audio pipeline to cuda / float16 GPU acceleration.

---

Core Features

• Push-to-Talk Privacy: Zero continuous microphone streaming. Audio capture triggers exclusively when you click 'Click to Speak', eliminating background surveillance.
• Agentic Code Execution: Translates user intent into verified Python automation with an integrated 30-second watchdog timer and automatic self-healing.
• Minimalist Control Center: Clean pywebview interface for managing runtime status, testing microphones, reviewing LLM output, configuring AI keys, and managing plugins.
• Mobile Web Remote: Access and control your assistant from any device on your local Wi-Fi network (Port 8765) with live status and quick action tiles.
• 3-Tier Security Guard: Strict permission gating (Safe Level 1 auto-execute, Sensitive Level 2 confirmation prompt, Critical Level 3 blocked).
• GitHub Integration: Authenticated repository status, issue/PR creation, and branch pushes via Windows Credential Manager.
• Built-in Plugins: Native controllers for Google Chrome, Brave Browser, WhatsApp, and GitHub.
• Dynamic Path Resolution: Dynamically resolves active OneDrive, Desktop, Downloads, and Documents paths.

---

Installation & Running from Source

# 1. Clone repository
git clone https://github.com/PRN-6/PRIVACY68.git
cd PRIVACY68

# 2. Create virtual environment with Python 3.11 or 3.12
python -m venv .venv
.\.venv\Scripts\Activate.ps1

# 3. Install dependencies
python -m pip install --upgrade pip
pip install -r requirements.txt .

# 4. Start PRIVACY68
python app.py

To open the dashboard directly:

python app.py --dashboard

---

Building the Windows Installer

Building a standalone distributable installer uses PyInstaller and Inno Setup 6:

.\.venv\Scripts\python.exe -m pip install -r requirements.txt .
.\build.bat

This compiles PRIVACY68.exe under dist\PRIVACY68\ and packages an Inno Setup installer under installer\PRIVACY68-Setup-v1.6.2.exe.

---

Unit Testing

Run the full automated test suite (87 unit tests covering agent brain failover, execution self-healing, command executor, security guard, GitHub integration, and dashboard bridge):

python -m unittest discover tests -v

---

Project Documentation

• Project Documentation Portal
• About & Architecture
• Plugin Architecture Guide
• Security Policy
• Contributing Guidelines
• License