Research Papers MCP: PubMed, arXiv, Crossref

An MCP server that gives Claude, ChatGPT, Gemini, Cursor and any other MCP client seven research tools over 180M+ scholarly works:

Tool	What the agent gets
search_papers	Search Crossref, the DOI registry of 180+ million journal articles, conference papers, books, preprints and reports from every field, by words, author, journal, years and type, by relevance, newest or most cited.
search_biomedical	Search the life-science literature of Europe PMC: every PubMed record, PubMed Central and preprints (bioRxiv, medRxiv, Research Square...), 45+ million records, with abstracts, citation counts and open access status.
search_arxiv	Search arXiv preprints (AI and computer science, physics, mathematics, statistics, quantitative biology, economics) by words, category, author and years.
get_paper	Everything about one paper: all authors with ORCID and affiliations, the whole abstract, keywords and MeSH terms, citation and reference counts, funders, license and links.
get_references	The reference list of a paper, in its order, with the DOI or PubMed ID of each cited work.
get_citations	The papers that cite a paper, newest or most cited first.
read_full_text	An open access article in full, as text with its sections and tables, when its license allows reuse.

Ask your agent things like "What has been published this year on GLP-1 receptor agonists and dementia? Summarize the evidence", "Find the most cited papers on microplastics in drinking water", "What are the newest arXiv papers on LLM agents?", "Who followed up on this CRISPR study, and did anyone contradict it?" or "Read the methods of this trial and list its limitations": it picks the tool and gets clean JSON back. Every result carries an id (DOI, PubMed ID or arXiv ID) that the other tools take. You pay only for results: a search that finds nothing, an unknown id or an article that cannot be read costs nothing.

Connect your agent

Easiest, with your Apify account (sign-in in the browser, no token to copy): add this remote MCP server URL to your client:

https://mcp.apify.com?tools=digital_influx/research-papers-mcp

• Claude (claude.ai, Claude Desktop): add a custom connector with that URL.
• ChatGPT: add it as a connector (MCP server URL) in developer mode.
• Cursor, VS Code, Windsurf, Gemini CLI and other clients: add a remote MCP server. With a token instead of the browser sign-in:

{
  "mcpServers": {
    "research-papers": {
      "url": "https://mcp.apify.com?tools=digital_influx/research-papers-mcp",
      "headers": { "Authorization": "Bearer YOUR_APIFY_TOKEN" }
    }
  }
}

(Gemini CLI writes httpUrl instead of url in ~/.gemini/settings.json.)

Directly, without the Apify MCP server: this Actor runs in Standby mode as a Streamable HTTP MCP server. Take its Standby URL from the Actor's API tab in Apify Console, add /mcp, and send your Apify token as Authorization: Bearer YOUR_APIFY_TOKEN.

A record of every call: each tool call your agent makes is also saved as a row of the Standby run's dataset (tool, arguments, a one-line summary and the result), so you can check later what it asked and got. Without an agent: a standard run calls one tool from the input (tool and arguments) and saves the result to the dataset, so the tools also work from the API, schedules and integrations (Make, n8n, Zapier).

The tools

search_papers

{ "query": "microplastics", "fromYear": 2020, "toYear": 2021, "type": "journal-article", "sort": "cited", "limit": 10 }

All arguments are optional, but give at least a query, an author or a journal: fromYear, toYear, type (journal-article, proceedings-article, posted-content for preprints, book-chapter, book, report, dissertation, dataset), onlyWithAbstract, sort (relevance, newest, cited) and limit (10 by default, at most 50). Returns title, first authors, year, journal, citation count, DOI and the abstract when the publisher deposited it (first 1,000 characters). Crossref matches words, not meaning: add an author, a journal or years, or sort by citations, to see the landmark papers first. Crossref also matches any word of an author's name, so with author the tool reads more results and keeps only the works with an author of that surname (and initial, when you give the first name): "Geoffrey Hinton" sorted by citations gives Deep learning, ImageNet classification with deep convolutional neural networks and Learning representations by back-propagating errors, without the papers of other Geoffreys.

Real result (2026-10-04): the most cited microplastics papers of 2020-2021 are Plasticenta: First evidence of microplastics in human placenta (Environment International, 3,218 citations) and Environmental exposure to microplastics: An overview on possible human health effects (2,495).

search_biomedical

{ "query": "GLP-1 receptor agonists AND dementia", "openAccessOnly": false, "sort": "relevance", "limit": 10 }

The query takes Europe PMC syntax: phrases in quotes, AND / OR / NOT, and fields such as TITLE:, ABSTRACT: or MESH:. Optional: author ("Doudna JA" or a surname), journal, fromYear, toYear, openAccessOnly, preprintsOnly, reviewsOnly, sort and limit. Each paper comes with its PubMed ID, PMC ID and DOI, journal, year, citation count, publication types, open access status, whether read_full_text can read it, and the abstract.

Real result (2026-10-04): 3,109 papers and preprints match GLP-1 receptor agonists AND dementia; one of them, a 2026 PLOS One cohort study, is CC BY and can be read in full.

search_arxiv

{ "query": "retrieval augmented generation", "category": "cs.CL", "sort": "newest", "limit": 10 }

Optional: query (words in titles and abstracts, phrases in quotes), category (an arXiv code: cs.LG, cs.CL, math.AP, q-bio.NC, hep-th...), author, fromYear, toYear, sort (relevance or newest) and limit. Returns title, authors, submission date, version, categories, the comments field ("Accepted to NeurIPS"), the abstract and links to the arXiv page and PDF.

get_paper

{ "id": "10.1038/nature14539" }

id is a DOI, a PubMed ID (23287718), a PubMed Central ID (PMC3795411), a Europe PMC preprint ID (PPR1314996) or an arXiv ID (1706.03762), bare or as a link. Crossref and Europe PMC are merged: title, all authors with ORCID and affiliations, journal, publisher, volume, issue and pages, the whole abstract, keywords, MeSH terms, citation count (Crossref and Europe PMC), reference count, funders and grants, license, open access status and links. arXiv papers come from DataCite, with the citation count of OpenCitations.

Real results (2026-10-04): Deep learning (LeCun, Bengio and Hinton, Nature 2015) has 77,915 citations in Crossref and 103 references; Attention Is All You Need is at version 7 (last revised 2023-08-02) with 24,956 citations counted by OpenCitations.

get_references

{ "id": "10.1038/nature14539", "limit": 25 }

The reference list in the paper's order: title, authors, year, journal, volume, pages, and the DOI or PubMed ID of each cited work, or the citation text when the publisher deposited only that (DOI-only references get their title from Crossref). From Crossref, where most publishers deposit open references, or Europe PMC. limit: 25 by default, at most 200.

get_citations

{ "id": "23287718", "sort": "cited", "limit": 20 }

The papers that cite a paper, from Europe PMC's citation network (life sciences: PubMed, PMC and preprints), newest (default) or cited first, with optional fromYear, toYear and openAccessOnly. For a paper outside Europe PMC it returns the citation count from Crossref or OpenCitations, free.

Real result (2026-10-04): 11,830 papers cite Cong et al. 2013, Multiplex genome engineering using CRISPR/Cas systems; the most cited of them is Genome engineering using the CRISPR-Cas9 system (Nature Protocols, 9,863 citations).

read_full_text

{ "id": "PMC13502699", "sections": ["methods", "results"], "maxCharacters": 30000 }

The full text of an open access article from Europe PMC, as plain text with its headings (## Methods, ### Study design), figure and table captions, and the tables row by row; plus the list of sections with their length, so the agent can ask for only the ones it needs. maxCharacters: 30,000 by default, at most 100,000. Only articles whose license allows reuse are returned (CC BY, CC BY-SA, CC0, public domain); for CC BY-NC, CC BY-ND and closed articles the tool gives the link to read it, free.

How much it costs

Pay per event, only for results:

Event	When	Price
paper	each paper a search returns, and each reference or citing paper listed	USD 0.002
paper-details	each paper read with get_paper	USD 0.01
full-text	each article read in full with read_full_text	USD 0.02

A search of 10 papers costs USD 0.02. Searches that find nothing, unknown ids, papers without an open reference list and articles that cannot be read in full cost nothing. Runs in Standby use 256 MB.

Good to know

• Respectful by design: an honest User-Agent (research-papers-mcp), only public APIs at the pace each one asks for (Crossref's public pool, one request per second; Europe PMC, 10 seconds between requests as its robots.txt asks, so a call that follows another one there can wait a few seconds), no proxies, no login. arXiv is read through DataCite, where arXiv registers every paper, because arXiv's own API is closed to robots.
• Licenses are respected. Bibliographic metadata are facts free to reuse; abstracts remain under the copyright of their authors and publishers and come as Crossref and Europe PMC distribute them. Full texts are returned only under licenses that allow reuse, with title, authors, journal, DOI and license, so the credit goes with the text.
• Sources: Crossref, Europe PMC (EMBL-EBI), DataCite and OpenCitations. Citation counts differ between sources because each counts the citing works it knows.
• More tools for agents by the same author: Company Research MCP, Job Search MCP, Marketing & SEO Research MCP and US & UK Public Records MCP.