MixTape
I have spent a long time collecting music as files, on a disk I own. Every tool that is actually good at suggesting what to play next lives inside a service that would rather I streamed something else, and that keeps a record of everything I listen to.
MixTape is the other way around. The library is mine and sits on my machine, and the model reading it is mine too — LM Studio on the same Mac, no account, no API key. Nothing about what I listen to leaves the building.
It reads a Doppler library directly, and that library is the entire world the model is allowed to work in.
The interesting constraint is that a language model asked about music will confidently name records you don’t have. So it never gets to name anything — each prompt carries a numbered list of songs I own and the model answers with numbers into it. It can’t invent a track that isn’t there.
Discover asks the model for themes of its own and fills nine at once. Create takes a description — long biking trip, focused deep work — and builds one playlist for it. More like this starts from a single track you name and finds what sits beside it.
Whatever comes out is written back into Doppler as a native playlist, or exported as a plain .m3u.
