Summary

Literature work stops being a search box and becomes a standing position on a field.

Most paper tools return a flat list; this one ranks it, then lets you walk outward from any result through 22M+ citation edges. The part worth the setup is what persists between sessions — a library that tunes the feed, watches evaluated daily against new papers, and find_gaps telling you what the field considers canonical that you never saved.

What it is

A research server over Scholar Feed's arXiv index — 600k+ CS/AI/ML papers, updated daily, each carrying an LLM summary and a novelty score. It combines semantic and keyword search with citation-graph traversal, a saved library, and daily watches.

What you get

Semantic search that finds conceptually related work even when the wording differs, with mode='keyword' switching to exact full-text matching and novelty_min filtering on the 0.0-1.0 novelty score (search_papers).
Paper lookup by arXiv ID, batching up to 50 papers in one call, with format='bibtex' for a citation entry (get_paper).
Citation graphs in either direction ranked by citing-paper rank_score (get_citations), foundational reading for a topic (get_field_orientation), and the canonical anchors for one paper's own sub-field (get_foundational_lineage).
Full text pulled from a paper's LaTeX source with a PDF fallback — either the results and experiments excerpt with table captions, or sections='all' for the whole paper (fetch_fulltext).
Researcher lookup by topic or name, and the co-authorship neighbourhood around them with papers_count and last_collab_year on every edge (find_author, co_author_graph).
A saved library and named, nestable collections that double as the personalization signal for the feed (save_paper, unsave_paper, like_paper, list_library, list_collections, create_collection, add_to_collection, remove_from_collection).
Standing watches evaluated daily against newly indexed papers, with preview_watch dry-running a filter before you commit to it and check_watches pulling new hits without advancing the digest watermark (create_watch, update_watch, delete_watch, list_watches).
Three questions plain search cannot answer: find_gaps names important work missing from a collection, ask_library synthesizes only over papers you saved with inline arXiv-ID citations, and check_drift reports whether a method has been superseded and by what.
A 768-dim Gemini embedding for any text, for HyDE-style retrieval against the same vector space (embed_text).

Requirements

Node.js 18 or newer, launched as `npx -y scholar-feed-mcp@latest` over stdio; `npx scholar-feed-mcp@latest init` writes the client config for you. SF_API_KEY is optional and sets the daily volume: 100 calls/day anonymous, 1,000/day with a free key, 10,000/day on Pro. The library, collection, watch and gap tools name the key as a requirement; search, citation traversal and full-text extraction run without it. embed_text and find_gaps are Pro-only and return a 403 pro_required otherwise, and ask_library is 1/month free then 200/day on Pro.

Setup effort

One command — npx scholar-feed-mcp@latest init