liliu-z/stashbase

StashBase is a desktop app that keeps a personal library of notes, research papers, and other reference material on your own computer, then makes that library searchable from AI chat tools. It runs on macOS through a Homebrew install, with a build-from-source path for Intel Mac and Windows users. The project is in early alpha and the source is licensed Apache 2.0. The pitch is that most AI tools treat your reading material as temporary context: you upload a file, paste a note, or share a chat link, and then it is gone. StashBase aims to be a long-lived layer instead, watching folders of files on disk and keeping them indexed continuously. New notes become searchable straight away, and the index is exposed through the Model Context Protocol, so Claude, ChatGPT, Codex, Cursor, and other MCP-aware clients can ask for relevant snippets when you mention @stashbase in a chat. A folder you point it at is called a space. Each space gets its own local index built with two open libraries called mfs and Milvus Lite, combining meaning-based search with plain keyword search. The first time a space is indexed you choose an embedding model: OpenAI's text-embedding-3-small for about two cents per million tokens, or a built-in offline model called bge-m3 that needs no API key. StashBase ships with a built-in terminal that has Claude Code and Codex pre-wired. Anything those agents write to disk is picked up by the index without a manual refresh. Reusable agent instructions placed in a space's skills folder are mirrored into the formats both CLIs expect. Renaming a note triggers a confirmation dialog that rewrites links pointing at the old path, and dropping in a PDF runs a pipeline that turns it into a readable HTML note. The author argues that HTML is a better long-term note format than Markdown once a language model is doing most of the writing, because HTML carries richer structure such as semantic sections, anchors, tables, and embedded media. Markdown stays supported for drafts alongside HTML.