About

I built Locus for myself first. For about five years I worked on the feedback loops that make AI models better — reviewing outputs, judging quality, the unglamorous work behind every model release you've used. It leaves you with a strange, specific instinct: you know exactly how much a model's answer depends on what it's actually shown, and how often the real bottleneck isn't the model at all — it's that the answer was sitting in a file on someone's machine the whole time, and nothing ever looked there.

The moment this became personal was ordinary: on a video call, switching between browser tabs and a terminal, trying to find a decision I knew I'd written down somewhere — a note, a doc, an old message — while the AI assistant I was also running had no way to see any of it. It could reason well. It just couldn't look. Everything useful I knew was scattered across folders, and every AI tool I used started from zero every time.

So I built the thing I actually wanted: point it at your files once, and the AI assistants you already use that support MCP — GitHub Copilot, Cursor, Claude Desktop, your favorite LLM's CLI — can search them and cite exactly where an answer came from. Not a new assistant to learn. The same ones you already use, finally able to see what you've already written.

Locus is local-first because that origin made the constraint obvious: your files, your decisions, your history — they stay yours. Your raw files are never uploaded to Locus. When your AI assistant searches, the matching excerpts go to that assistant. If you turn on hosted features, content passes through our relay on the way, and we don't store it.