DeepGit

Why we included this project

DeepGit is built for people who have ever typed something specific into GitHub search and gotten back a pile of keyword matches that miss the point. It keeps your full request as requirements rather than compressing it into tags, searches across keyword, topic, and semantic angles, weighs each candidate on real evidence, and only reads actual source code when the top picks are too close to call. The cost control is what makes it usable day to day: a typical query runs about two LLM calls, with code reading reserved for genuinely contested results. It also enforces constraints as hard rules, so a C++ request will not surface a C library, and its local semantic index (LanceDB, CPU-only) lets it remember repos whose descriptions never mention the obvious words. A built-in MCP server means you can hand the whole research job to your own assistant instead of running searches by hand.