Citation verification, bibliography auditing for retractions and dead links, and a recommendation auditor that looks for self-promotion are unusual things for a search server to carry, and they are the reason to choose this one. The session model is the other half: research accumulates across steps and exports with per-step provenance, so the report says where each claim came from.
A research server that pairs general web search and full-page reading with specialist indexes — academic, legal, patent, clinical, economic, biomedical — and a set of tools for auditing what came back.
- Web search that can be narrowed to only the sources you trust, plus a combined search-and-read that scores results for reliability before fetching them
- Full-document reading rather than snippets: web pages, PDFs, Word documents, slide decks, YouTube transcripts and Hacker News threads, with a raw mode for verbatim source
- Academic work handled properly — papers with real DOIs, authors and citation counts, full text fetched in one call, and a citation graph walked in both directions with influence signals
- Specialist indexes for the questions general search answers badly: US court opinions and dockets, patents across offices, ClinicalTrials.gov registrations, World Bank, OECD, Eurostat and FRED economic series, and the Monarch biomedical knowledge graph
- Verification as first-class work: check whether a single citation exists, matches a real record, is retracted or dead; audit a whole bibliography in one pass; and audit an AI-generated recommendation list for self-promotion, conflicts and dead links — evidence rather than a verdict
- Sources preserved by capturing a fresh Wayback Machine snapshot, so a cited page stays checkable after it changes
- Multi-step research that keeps state: a session that builds on what it already found, recoverable after a context loss, exportable as a Markdown or JSON report with per-step provenance
- Bibliographies formatted as APA, MLA, BibTeX, RIS or CSL-JSON for Zotero, EndNote and Mendeley
- Company work — brand identity pulled from a domain, and OSINT reconnaissance from certificate transparency logs, Wayback history and derived subdomains, each phase failing soft
- Prompt templates that walk a model through comprehensive research, fact-checking and competitive analysis
No key to start: DuckDuckGo is the built-in zero-config search backend. Adding any one provider raises result quality and unlocks image and news search — Google Programmable Search needs GOOGLE_CUSTOM_SEARCH_API_KEY with GOOGLE_CUSTOM_SEARCH_ID, and Brave, among others, is interchangeable with it; configure more than one and the server fails over between them. A few tools activate only with their own configuration, such as FRED for US macro series and a Brave key for local search. Run it with uvx, from the published container image, or as a bundle for Claude Desktop.
One command — uvx web-researcher-mcp