scholar-workflow

Working effectively with scholarly sources

When the user asks about academic literature, prior art, citations, patents, books, or standards, use the scholar-mcp tools as described below. The server covers four source domains: papers, patents, books, and standards. Each domain has its own tools, but they cross-reference naturally — a paper may cite a patent, a patent's NPL section may reference a paper, and a paper with an ISBN is automatically enriched with book metadata.

Choosing the right search entry point

User intent	Tool	Key parameters
Find papers on a topic	search_papers	query, year_start/year_end, fields_of_study, venue, min_citations, sort
Look up a specific paper	get_paper	identifier — DOI, S2 ID, ARXIV:id, ACM:id, PMID:id
Find an author's work	get_author	Numeric S2 author ID for profile, or name string for disambiguation
Find patents	search_patents	query, cpc_classification, applicant, inventor, jurisdiction, date filters
Look up a specific patent	get_patent	patent_number, sections (biblio, claims, description, family, legal, citations)
Find books	search_books	Prefer title/author over query — dedicated indexes give better recall
Look up a book by ISBN	get_book	identifier — ISBN-10, ISBN-13, OL work ID, or edition ID
Find a standard	search_standards	query, optional body filter (NIST, IETF, W3C, ETSI)
Normalise a messy citation	resolve_standard_identifier	raw — e.g. "rfc9000", "nist 800-53", "wcag 2.2"
Resolve a mixed list of IDs	batch_resolve	identifiers — mix of DOIs, S2 IDs, patent numbers, ISBN: prefixed ISBNs

Domain-sticky traversal: a paper's references mostly point to other papers; a patent's family members are other patents; a standard often cites other standards. Follow the domain naturally — don't switch tools mid-chain unless the data leads you there (e.g. a patent's NPL citations resolve to papers).

Literature search strategies

Broad survey

1. search_papers with a broad query. Use sort="citations" and min_citations to surface influential work quickly.
2. Pick the most relevant hit and call get_citations to see who built on it.
3. Call get_references on the same paper to see what it builds on.
4. If the topic spans disciplines, use fields_of_study to split searches (e.g. search once for "Computer Science", once for "Medicine").

Targeted lookup

• When the user gives a DOI, arXiv ID, or other identifier, go straight to get_paper. Don't search first.
• For author-centric queries, get_author with a name returns up to 5 disambiguation candidates. Ask the user to pick before fetching publications — don't guess.

Patent prior art

1. search_patents with keywords, CPC codes, or applicant names.
2. get_patent with sections=["biblio", "citations"] — the citations section includes NPL references automatically resolved to S2 papers.
3. get_citing_patents finds patents that cite a given paper (coverage is incomplete — EPO OPS does not capture all citations).

Standards lookup

• resolve_standard_identifier handles messy input ("rfc 9000", "NIST SP 800-53 rev5") and normalises to canonical form. Use it first when the user gives an informal name.
• get_standard with fetch_full_text=true downloads and converts the standard via docling (requires SCHOLAR_MCP_DOCLING_URL). This may block until conversion completes; it only returns a task ID if rate-limited.

Citation graph traversal

Forward and backward citations

• get_citations — papers that cite a given paper (forward). Supports year, field, and minimum citation count filters.
• get_references — papers cited by a given paper (backward).
• Both return paginated results. Use offset for additional pages.

BFS graph expansion

get_citation_graph does breadth-first expansion from 1–10 seed papers.

• depth 1 is usually sufficient for a neighbourhood view. Depth 2–3 expands rapidly — always set max_nodes to cap growth.
• direction="both" finds the densest cluster but doubles API calls per hop. Use "citations" or "references" when the user has a clear direction.
• Combine with year_start/year_end and min_citations to keep results focused. When filters are active, the tool fetches more candidates per node to compensate for filtering losses.

Bridge papers

find_bridge_papers finds the shortest citation path between two papers. Useful for connecting seemingly unrelated work. Searches up to max_depth hops (default 4). Use direction="both" for best coverage.

Recommendations

recommend_papers takes 1–5 positive paper IDs and optional negative IDs. The positive examples define the topic; negatives steer away from unwanted areas. For best results, pick positive examples that span the desired topic rather than clustering around one sub-area.

recommend_books takes a subject string (e.g. "machine learning", "algorithms") and returns popular books from Open Library sorted by edition count.

Citation generation

generate_citations produces BibTeX, CSL-JSON, or RIS for up to 100 papers.

• enrich=true (the default) adds OpenAlex venue metadata for more complete BibTeX entries (journal names, volumes, pages).
• Unresolved paper IDs are included in the output with a "not found" note rather than silently dropped.

Enrichment

• Book enrichment is automatic: any paper with an ISBN in externalIds gets a book_metadata field (publisher, edition, cover URL, subjects) from Open Library. No extra tool call needed.
• enrich_paper adds OpenAlex fields (affiliations, funders, OA status, concepts) on demand. Useful when the user needs institutional or funding context beyond what S2 provides.

Full-text retrieval and PDF conversion

PDF tools are write-tagged and hidden in read-only mode (the default). The user must set SCHOLAR_MCP_READ_ONLY=false and have a running docling-serve instance (SCHOLAR_MCP_DOCLING_URL) for Markdown conversion.

Papers

• fetch_paper_pdf downloads the PDF with automatic fallback: S2 open-access → ArXiv → PubMed Central → Unpaywall.
• fetch_and_convert does fetch + convert in one call — usually what users want.
• fetch_pdf_by_url handles arbitrary PDF URLs (e.g. an author's website).
• convert_pdf_to_markdown converts a local PDF file to Markdown — use this when you already have the PDF on disk (e.g. a paywalled paper obtained manually). Accepts an absolute file_path.

Patents

• fetch_patent_pdf downloads via authenticated EPO OPS. Not all patents have PDFs available (older patents, some WO publications).

Standards

• get_standard with fetch_full_text=true fetches and converts in one call.

VLM enrichment

Pass use_vlm=true to fetch_and_convert, fetch_pdf_by_url, fetch_patent_pdf, or convert_pdf_to_markdown for better formula and figure extraction. (get_standard uses docling internally but does not expose the VLM flag.) Requires SCHOLAR_MCP_VLM_API_URL and SCHOLAR_MCP_VLM_API_KEY. VLM and standard conversions are cached separately — switching modes never overwrites previous results.

Handling queued operations

PDF downloads (fetch_paper_pdf), patent PDF fetches (fetch_patent_pdf), full pipeline runs (fetch_and_convert), URL-based fetches (fetch_pdf_by_url), local file conversions (convert_pdf_to_markdown), and any tool that hits a rate limit return a task ID when work is submitted to the background queue. Cache hits return immediately without queuing. Example queued response:

{"queued": true, "task_id": "abc123", "tool": "..."}

Poll with get_task_result(task_id="abc123"). The response includes status, elapsed_seconds, and a hint while running.

list_tasks shows all active tasks. Don't poll in a tight loop — tasks typically complete in 10–30 seconds (PDF download) or 1–5 minutes (PDF conversion with VLM).

Do not

• Do not call search_papers when the user gives a specific identifier — use get_paper directly.
• Do not call get_author with a name and then silently pick the first candidate — ask the user to disambiguate.
• Do not set depth > 2 in get_citation_graph without also setting a tight max_nodes cap — the graph expands exponentially.
• Do not use fetch_pdf_by_url for EPO OPS patent URLs — they require authentication. Use fetch_patent_pdf instead.
• Do not forget batch_resolve when the user gives a mixed list of identifiers — it routes each to the correct backend automatically.