Glass by Pickle is an open-source, local-first desktop application designed to act as a digital mind extension by passively observing your screen and audio inputs. Built with Electron, Next.js, and Tailwind CSS, it intelligently synthesizes your interactions—like meetings or screen activity—into structured notes, action items, and answers without recording or sharing your data. Unlike traditional screen recorders or cloud-based assistants, Glass runs entirely on-device and is designed to be invisible: it doesn’t appear in screenshots or screen recordings. This makes it ideal for privacy-conscious professionals who want AI-powered context without surveillance.
The app is a fork of CheatingDaddy and has been enhanced with support for multiple LLMs (OpenAI, Gemini, Claude), local models via Ollama and Whisper, Windows beta support, and real-time audio separation using Rust. Its mission aligns with Pickle’s broader goal of building a living digital clone by transforming daily data into scalable personal knowledge.
What You Get
- Real-time meeting summaries - Automatically generates structured notes and action items during meetings by analyzing audio and screen activity, helping you focus without manual note-taking.
- Context-aware AI queries - Ask questions using Ctrl/Cmd + Enter to get answers based on your entire screen and audio history up to that moment, not just isolated prompts.
- Multi-provider LLM support - Use your own OpenAI, Gemini, or Claude API keys—or switch to local LLMs via Ollama and Whisper for full data privacy.
- Invisible operation - The app does not appear in screenshots, screen recordings, or the system dock; no background recording or data sharing occurs by default.
- Local-first architecture - All processing happens on-device; optional Firebase storage is available but not required, ensuring user data remains under your control.
- Cross-platform support - Runs on macOS and Windows (beta), with audio separation via Rust for clearer mic vs. system audio input.
Common Use Cases
- Building a private AI meeting assistant - Professionals in consulting or engineering use Glass to auto-generate post-meeting summaries without uploading sensitive conversations to cloud services.
- Developers in regulated industries - Teams in healthcare or finance use Glass to capture meeting insights without violating data governance policies, thanks to its on-device processing.
- Problem: Forgetting key action items during meetings → Solution: Glass listens, analyzes, and surfaces tasks in real time without interrupting the flow.
- Team workflows with distributed developers - Remote teams rely on Glass to maintain context across async meetings, using its local-first design to avoid reliance on unstable cloud APIs.
Under The Hood
This project is a cross-platform desktop AI assistant that combines an Electron-based desktop application with a web frontend built using Next.js and React. It integrates cloud and local data storage, supports multi-modal AI interactions, and enables real-time communication between desktop and web components through a hybrid architecture.
Architecture
This system follows a layered monolithic design that decouples the Electron desktop app from its web frontend, enabling modular feature development.
- The architecture uses a layered pattern where the Electron main process handles system-level operations and the web frontend manages UI and user interactions via a Node.js backend.
- Modules are organized by feature areas such as AI interaction, audio processing, and user management with dedicated services and repositories.
- Design patterns like service locator, adapter pattern, and event-driven communication are employed to support modularity and decoupling.
Tech Stack
The project leverages a modern web and desktop tech stack to support AI-powered features and cross-platform functionality.
- Built primarily with JavaScript and TypeScript, using Electron for desktop app development and Next.js for the web interface.
- Integrates Firebase for authentication and real-time database, Express.js for backend routing, and various AI providers such as OpenAI and Anthropic.
- Utilizes esbuild for building the renderer, Electron-builder for packaging, and Tailwind CSS for styling the frontend.
Code Quality
The codebase reflects a mixed level of quality with some structured patterns but areas needing improvement in consistency and testing.
- Code linting and formatting are configured with ESLint and Prettier, indicating a commitment to code style and maintainability.
- Error handling is present but not uniformly applied across all modules, leading to potential inconsistencies in robustness.
- Technical debt is visible in areas like circular dependencies and limited test coverage, affecting long-term maintainability.
What Makes It Unique
The project stands out through its innovative hybrid architecture and extensible service model for AI and data handling.
- Introduces a novel bridge pattern that decouples Electron’s main and renderer processes while enabling seamless inter-module communication.
- Offers a unified provider abstraction for integrating local and cloud AI models, supporting flexible backend switching.
- Features an extensible repository structure that allows for encryption-aware persistence and easy data source swapping.