jcarterjohnson/constellate

Constellate takes your entire chat history with AI assistants, across every provider and model you have used, and turns it into an explorable 3D map. Conversations that cover similar topics are pulled close together, so where a conversation sits in the map tells you how related it is to the others. A second view arranges conversations by time on one axis and their two strongest topic differences on the others. The project ships as two separate apps that share the same underlying engine and file format. The first is a web app that is just a single HTML file with a copy of the Three.js graphics library included, so you open it in a browser and there is nothing to install. The second is a native Mac app built with SwiftUI and rendered with SceneKit and Metal, which does everything the web app does plus some extra features when the hardware supports them. A core idea behind the project is privacy. Everything runs on your own computer: importing your chat exports, finding similarities between conversations, calculating the layout, and drawing the 3D scene. Nothing is uploaded anywhere and there are no accounts or servers involved. You bring your own exported chat history from services like ChatGPT, Claude, or Gemini, or you can drop in markdown files. The repository itself contains no real personal conversations, only a small invented demo dataset you can load to try the app out. The web app can be opened by double-clicking the file, and it supports drag-and-drop importing, with an optional folder-watching feature in Chrome or Edge that automatically picks up new export files. The Mac app has to be built from source code using Apple's Xcode tools, and it requires an Apple silicon Mac. It has optional AI features, like generating plain-language names for topic clusters, that only work if a setting called Apple Intelligence is turned on, without it the app still works for browsing, searching, and importing, it just skips those specific features and says why.