The handle discipline runs through the whole tool set — search returns summaries, excerpts preview before you commit, and a section can be fetched with its ancestors and child summaries so a thin heading does not send you back to reading the file. The other half of the surface is upkeep rather than reading: link integrity, orphans, drift against disk, duplicate clusters, and a health radar with an A-F grade. The saving depends on the shape of your corpus — a 40-line file with one heading saves nothing, a large reference site saves a great deal — and the code-side tools do nothing unless the companion code indexer is installed alongside.
A documentation index you run locally. It parses folders and GitHub doc trees into sections by heading hierarchy and serves those sections by handle, with 64 tools covering retrieval, link and freshness audits, OpenAPI lookups and index upkeep.
- Markdown, MDX, reStructuredText, AsciiDoc, notebooks, HTML, OpenAPI, JSON and XML parsed per format, with PDF, DOCX, PPTX and EPUB added by the optional office extra for local indexing
- Retrieval that stops at the size you asked for: search returns summaries, an excerpt previews 500 bytes by default, and full content is byte-range-read from the original file
- Hybrid ranking that fuses BM25 with embeddings when the index was built with them, plus a title-only match for pure navigation
- Structure audits over the whole corpus: cross-references that no longer resolve, sections nothing links to, backlinks into a document, wiki pages whose declared sources have changed on disk
- Rewrite safety before you touch anything: transitive blast radius over the inbound reference graph, tutorial-path membership, and a single safe-to-delete verdict with ranked blockers
- OpenAPI lookups once a spec is indexed — operations by path glob, method or tag, the schema reference graph from a root schema, and which operations pull a given schema in
- A glossary extracted at index time, and a bridge to a jcodemunch code index reporting which symbols have documentation and which do not
- Freshness disclosed rather than assumed: content reads report which source layer answered and whether the section has drifted from disk
Python 3.10 or newer and documentation to point it at, on disk or in a GitHub repo; uv tool install jdocmunch-mcp followed by its init writes the client entries for you. Everything runs locally — indexes live under your home directory — and the only default network call is an anonymous savings counter, switched off with JDOCMUNCH_SHARE_SAVINGS=0. Hybrid search needs an embedding provider configured, and stays lexical without one. Licensing is the thing to settle first: free for personal and non-commercial use, and commercial use requires a paid licence, priced per seat tier by the vendor.
One command — uv tool install jdocmunch-mcp