Millfolio is a Mojo application that answers questions about your own documents while keeping the data on your machine. It uses two models with very different roles:
• A frontier model writes a small Mojo program to answer your question — but it only ever sees an aliased schema of your vault (file_0, col_2, …), never your filenames or contents.
• That program runs locally in a sandbox, using a small on-device model to actually read your files and produce the answer.
So the capable model does the reasoning and codegen, and only the local model ever touches your data.
The backend is Mojo end to end — on-device inference on Apple Silicon (Metal), a LanceDB vector index, and PDF/CSV/Markdown readers, all as Mojo libraries. It builds on @Ehsan ’s “How I built a pure Mojo app and 10 libraries with AI agents” and uses their flare HTTP library directly — thank you!
• Site & how it works: https://millfolio.app
• Live demo: https://demo.millfolio.app — the frontier-model codegen is replayed from a recorded cache, but the on-device inference and sandboxed execution are real and live (on a Mac mini).
• Code & discussion: millfolio · GitHub
This this an early experiment, feedback is appreciated 