Search, outline, navigate, read — four calls you can show someone, in a domain where 'the model found it somewhere' is not an acceptable answer. Keeping the working set small is the other half: the agent takes the section it needs rather than a window full of chunks. Note that the skills carry as much of the value as the tools; without them an agent will use search_documents like a search box and stop there.
Document retrieval over markdown, CSV and JSONL with no vector database, no embeddings and no model calls at index time. Search returns positions in a navigable tree; the agent then reads the outline, opens a node, or takes a section with its descendants.
- search_documents — BM25 with facet filters from your frontmatter, expanded through a glossary you supply
- get_tree for the table of contents: headings, word counts and summaries
- navigate_tree for a section plus everything under it in one call
- get_node_content for the full text of one section by node id
- lookup_row for exact-key retrieval from CSV and JSONL rows
- find_similar, draft_wiki_entry and write_wiki_entry when writing is switched on — with path containment, schema validation, duplicate guards and a dry run
- Bundled skills that teach the drill-down policy: search, outline, navigate, retrieve
- An init command that scaffolds a wiki layout, writes the client config and installs a lint hook
DOCS_ROOT pointing at your markdown, and that is the minimum. DOCS_GLOB widens it to CSV and JSONL, DOCS_ROOTS handles several weighted collections, and GLOSSARY_PATH supplies the query-expansion terms. Writing is off unless WIKI_WRITE is set. Runs over stdio by default; there is an HTTP mode for teams and hosted agents.
One command — bunx doctree-mcp init