Invariant testing is a subject where a matched paragraph is often actively misleading — the useful part is the setup and the caveat around it. Tools that return the complete chapter or post let a model read the argument rather than reassemble it from fragments. Searching three sources in one call is the other practical piece, because this corpus is genuinely split across a docs site, a book and a newsletter, and knowing which one holds an answer is not something a reader should have to know first. refresh_cache is a small thing that says a lot: documentation caching is a normal cause of a confidently outdated answer, and having a way to clear it is better than having no explanation for why the answer is old.
A documentation retrieval client over three Recon properties: getrecon.xyz with its blog, glossary, comparisons and tool listings; the Recon Book of chapters and technical concepts; and the Recon Substack archive.
- search_all queries every source at once, with search_site, search_book and search_substack when the source matters.
- Retrieval returns full documents rather than excerpts — get_book_chapter, get_blog_post and get_substack_post each return the complete content with metadata.
- get_book_concept covers technical concepts directly, and search_glossary returns the top matching terms for vocabulary questions.
- get_comparison returns both entities with their strengths, which is a shape a plain search result cannot give you.
- list_book_chapters groups by category and list_substack_posts sorts by date, so the corpus can be browsed rather than only queried.
- list_tools enumerates the developer tools documented on the site.
- refresh_cache forces a reload from all three sources, so a stale answer has a remedy that does not involve restarting anything.
Nothing to supply — the sources are public.
One command — npx @recon-fuzz-mcp/knowledge