The shallow one answers where is this file instantly; the deep one answers what is in it and is built only when asked. That split, plus literal-by-default search, is what keeps a large repository navigable without turning every question into a full scan.
A code index for assistants. Point it at a project and it builds a shallow file index for quick discovery, with a deeper symbol index built on request when you need structure rather than location. Search is literal by default, with regex and fuzzy matching available and results paginated.
- A project initialised by path and indexed, with the index rebuilt on demand after files change
- A deep symbol index built separately, because symbol-level analysis costs more than file discovery and is not always needed
- Search that is literal unless you ask otherwise, with regex, fuzzy matching, file-type filtering and paging — ten results a page by default
- Files located by glob pattern rather than by guessing at paths
- A file summarised into its structure, functions, imports and complexity, and a named symbol's body pulled out on its own
- A file watcher that can keep the index fresh automatically, with its status and settings both readable and changeable
- The search backend re-detected on request, so installing ripgrep or ugrep is picked up without a restart
- The index storage directory created, checked, and cleared when you want to start over
Python 3.10 or newer with uv; launched over stdio as uvx code-index-mcp, or installed from PyPI with pip. It indexes a local directory, set with the project-path tool or passed as an argument at startup so the server comes up already pointed at the repository. Regex search needs a native search tool such as ripgrep or ugrep on the machine — the built-in fallback is literal only — and symbol summaries need the deep index built first. On Windows, uvx wants the standard profile directories present in the server's environment.
One command — uvx code-index-mcp