mediumbirb

A personal AI desktop assistant for macOS with context awareness, meeting recording, and multi-model comparison.

Built with Electron, React 19, TypeScript, and a native Swift helper for reading your active application context via macOS Accessibility APIs.

• Context-Aware AI Chat — Knows what you're working on by reading your active app, browser tabs, and window content via native macOS Accessibility APIs
• Multi-Model Support — Access Claude, GPT-4o, Gemini, DeepSeek, Llama, Grok, and more through OpenRouter
• Arena Mode — Compare 2-4 AI model responses side by side with voting
• Meeting Assistant — Auto-detect meetings (Zoom, Google Meet, Teams), record audio, transcribe via Whisper, generate AI summaries
• AI Journal — Auto-generated daily journal entries from your activity and context
• Quick Access Overlay — Global hotkey (Cmd+Shift+Space) slides in an AI panel from the screen edge
• Privacy Controls — Exclude specific apps, websites, or entire categories (banking, health, etc.) from context capture
• MCP Integration — Connect any Model Context Protocol server for extensible tool support (Claude Desktop-compatible config)
• Dark & Light Mode — Clean minimal design with system preference detection

MediumBirb combines three systems to create a context-aware AI assistant:

1. Screen Context Reading (ContextKit) A native Swift CLI binary runs alongside the Electron app, communicating over JSON-RPC 2.0 on stdin/stdout. It uses macOS Accessibility APIs (AXUIElement) to read the UI hierarchy of whatever app is in the foreground — extracting window titles, text content, and UI element values. For browsers (Chrome, Safari, Arc), it reads the active tab URL via AppleScript. This context is fed into AI conversations so the model knows what you're looking at.

2. AI Chat with Streaming Chat messages are sent to OpenRouter's API, which provides access to every major AI model (Claude,