Peeky: A Privacy-First AI Desktop Assistant for Windows with Offline Voice and Visual Interaction

Peeky is a desktop AI assistant designed specifically for Windows, focusing on privacy-first and fully offline operation. All data processing is done locally, supporting multi-modal interaction methods such as voice dialogue, screenshot analysis, camera visual recognition, and clipboard processing—no need to worry about privacy leaks. Its core philosophy is "See, Think, Help", and it runs local open-source large language models via Ollama to provide users with secure and convenient AI services.

Unlike traditional cloud-based AI assistants, Peeky keeps all computing processes on the local machine and runs open-source large language models via Ollama to achieve a truly privacy-first design. Its core philosophy is "See, Think, Help". Users can interact in multiple ways, and sensitive data never leaves the device, addressing the privacy concerns of cloud-based AI assistants.

Peeky supports multiple interaction features:

1. Voice dialogue: Use Google Speech API + Edge TTS when online; switch to faster-whisper + Windows SAPI when offline;
2. Screen capture: Drag to select an area and ask questions, with local multi-modal models analyzing the content;
3. Camera vision: Call the camera to take photos and recognize content;
4. Clipboard processing: One-click analysis of clipboard text;
5. Video Coach: Capture a baseline image, provide voice guidance, and verify task completion.

Peeky's tech stack ensures offline operation:

Function Stage	Online Solution	Offline Solution
Audio Capture	ffmpeg+DirectShow	ffmpeg+DirectShow
Speech Recognition	Google Speech API	faster-whisper(base)
Inference Calculation	Ollama Local Run	Ollama Local Run
Speech Synthesis	edge-tts(Aria)	pyttsx3+SAPI(Zira)
The system first detects the network; if there's no network, it automatically skips online services to avoid lag.

Peeky's privacy protection measures:

1. All model inference is completed locally; prompts, images, and responses do not leave the device;
2. Online services are only used when connected to the internet and with user permission; Google Speech only receives audio, and Edge TTS only receives text;
3. Interaction history is stored locally in memory.json, which users can delete or clear at any time.

Peeky is suitable for various scenarios:

• Screen capture: Explain interfaces, analyze charts, guide software operations;
• Video Coach: Teaching demonstrations, device maintenance, software training;
• Users with privacy needs: Fully offline to ensure data security;
• Text processing: Quick analysis, summary, and translation of clipboard content.

Installation requirements: Windows10/11, Python3.10+, local Ollama service, microphone permissions, 8GB disk space. The gemma4:e4b multi-modal model is recommended, and the faster-whisper base model is automatically downloaded on first launch. Summary: Peeky returns privacy control to users and is an important direction for privacy-first AI tools. With the improvement of local models in the future, it will become more important in the field of personal computing.