localhost6767

Most assistants start every conversation knowing nothing about what you were just doing. localhost6767 is the backend for one that does: a desktop overlay that knows which window is in front, what you just copied, and what it has seen you work on before.

Collecting context

A collector samples the machine — the frontmost application and its window title (read through System Events on macOS) and the clipboard — and turns them into a context record. A short TTL cache drops anything seen within the last fifteen seconds, so copying the same thing twice, or sitting on one window, does not flood the memory with duplicates.

Deciding what matters

Not everything on screen should be remembered, and some of it must never be. An intent step classifies each context before anything is stored: code, terminal errors and stack traces, technical terms, articles and docs, video and music, shopping. And it screens for secrets first — anything that looks like a password, OTP, token, API key or private key is refused outright.

Memory, ranked

Useful context goes to a persistent memory layer (Supermemory). When a question comes in, related memories are pulled back and ranked before they reach the model: results below a similarity floor are dropped, near-duplicates are collapsed to the best-scoring copy, and only the top few survive. The answer comes from OpenAI with that context in the prompt, and goes back to the overlay over a WebSocket.

active window + clipboard → collector → intent (and secret screen)
    → memory ── question → rank + dedupe → OpenAI → WebSocket → overlay

Built at a hackathon, in Bun and Express.