hvardhan878/ghostwork

Ghostwork is a Mac application that watches everything you do on your computer and gradually takes over repetitive tasks without you telling it to. It runs in the background, silently recording your clicks, keystrokes, and app activity through a separate tool called Screenpipe. Over time it learns the patterns in how you work and starts automating the parts that repeat. The memory system has four layers that build on each other. The first layer holds what is currently on your screen. The second layer records raw events every two minutes: which apps you switched to, what you typed, where you clicked. Each night while you are not working, Ghostwork runs what the README calls a sleep cycle: it analyses the recent raw events, identifies patterns like "whenever I finish a LinkedIn search I export the results to my CRM", and compiles frequently repeated sequences into stored skills that can be replayed automatically. Automation happens on a trust system. New automations start in supervised mode, where Ghostwork runs the action and shows you a notification with an undo shortcut. After you have accepted the same automation at least five times without rejecting it, it switches to running silently. Actions that affect things outside your computer, like sending an email or submitting a form, always require a single tap of approval before they go through. When it executes actions, Ghostwork first tries the macOS accessibility layer to click and navigate natively. If that does not work it falls back to taking a screenshot and asking a vision-capable AI model to figure out where to click. If a multi-step action fails partway through, it automatically undoes the steps it already completed. All data stays on your machine. There is no cloud sync, no telemetry, and personally identifiable information like email addresses and card numbers is stripped before anything gets sent to an AI model. Raw events are deleted after 90 days. It is built with TypeScript and Electron. Setup requires macOS 12 or later, Screenpipe installed and running, Node.js 20 or newer, and an API key for an AI provider. The project is GPL-3.0 licensed.