Back

notchLM

A macOS menu bar app for Computer Use Agents. Uses vision models to analyze the screen and execute mouse, keyboard, window management, and shell commands via natural language. Includes a UI grounding system, Computer Use Agent harness, and ScreenSpot-Pro benchmark suite.

SwiftSwiftUIOpenRouterVision ModelsComputer Use
0:00
0:00
The agent analyzes screen content using local and/or cloud vision models, then executes actions like area highlighting (Chat Mode), mouse clicks, keyboard input, and app launching with native macOS APIs. (Consistent agent responses in chat window are WIP)
Model Composition separates reasoning from UI grounding. Larger models handle task planning while smaller dedicated grounding models locate UI elements. Both grounding and planning models have the option to be run locally or in the cloud, but typically only the grounding model is run locally. API keys are stored in the macOS Keychain.
Model Composition separates reasoning from UI grounding. Larger models handle task planning while smaller dedicated grounding models locate UI elements. Both grounding and planning models have the option to be run locally or in the cloud, but typically only the grounding model is run locally. API keys are stored in the macOS Keychain.
0:00
0:00
The toolbar provides quick access to key functions such as starting a new conversation, selecting your agent mode (Chat Only, Safe Mode, Full Agent), toggling screen context, accessing app settings, and closing the notchLM interface.
Runs ScreenSpot-Pro benchmark subsets (Android Studio, Excel, MATLAB, etc.) to measure grounding accuracy. Supports local dataset folders and compares notchLM's custom harness against the official evaluation protocol.
Runs ScreenSpot-Pro benchmark subsets (Android Studio, Excel, MATLAB, etc.) to measure grounding accuracy. Supports local dataset folders and compares notchLM's custom harness against the official evaluation protocol.
Austin Heiss