kytmanov/synto

Synto is a tool that turns a folder of plain Markdown notes into a personal wiki, where the work of writing the wiki pages is done by a local AI model running on your own computer. The idea, which the author credits to a description by Andrej Karpathy, is that your raw notes are source material rather than the finished article. You drop in notes, and Synto identifies the concepts in them, then writes one cross-linked article per concept. When the same concept shows up in several notes, those notes feed into the same article instead of creating duplicates. Everything stays on your machine unless you choose to share it. The pipeline has four stages: import, ingest, compile, and export. It uses two different local models. A smaller model around 4 billion parameters reads each note and extracts the list of concepts and summaries. A larger model around 14 billion or more then writes the actual articles. The author explains this split as a way to keep things fast and cheap on normal consumer hardware, since concept extraction is easier than writing coherent linked prose. After running synto init, you get a vault layout with raw for your notes, wiki for the published articles (in Obsidian-compatible Markdown with wikilinks), a drafts folder for articles waiting for review, a SQLite database to track state, and a pack folder that exports the wiki in a form Claude, Cursor, or other file-aware AI agents can read. There is no vector database and no embeddings; queries are routed through an INDEX.json file. Several mechanisms make the workflow safer. Compilation is incremental, so changing one note only rebuilds the articles tied to it. You can reject a draft and write a reason, which gets fed into the next prompt. Drafts get a confidence score from 0 to 1. Articles you edit by hand are tracked by a SHA-256 hash, so a recompile never overwrites your edits. Source documents carry a type like paper, textbook, or api_docs, and the ingest prompt adapts accordingly. The project is a Python tool published to PyPI, has 20 stars, and succeeds an earlier Obsidian plugin by the same author.