Crossref Papers MCP Server

An MCP server that connects Claude to Crossref, the largest scholarly metadata registry with 150M+ works. Search papers, explore citations, find author publications, and trace reference networks — all from within Claude.

No API keys required. The Crossref API is free and open. Optionally set CROSSREF_MAILTO for faster rate limits via Crossref's polite pool.

Features

• Paper search — search across 150M+ scholarly works with filters by year, type, and sort by relevance or citation count
• Citation exploration — find all papers that cite a given work, sorted by impact
• Author lookup — get an author's publications with total citation count and most-cited work
• Reference lists — retrieve and resolve a paper's bibliography with DOI-linked metadata
• Topic discovery — find the most influential papers in any subject area

Quick Start

Prerequisites

• Python 3.10+
• uv (recommended) or pip

Installation

git clone https://github.com/arterorx/crossref-mcp.git
cd crossref-mcp
uv sync

Configure in Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "crossref": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/crossref-mcp", "run", "crossref-mcp"]
    }
  }
}

Configure in Claude Code

claude mcp add crossref -- uv --directory /absolute/path/to/crossref-mcp run crossref-mcp

Polite pool (optional)

Set CROSSREF_MAILTO to get priority access to the Crossref API:

export CROSSREF_MAILTO="your@email.com"

Restart Claude after adding the configuration.

Tools Reference

search_papers

Search for scientific papers by keyword with filters.

Parameter	Type	Default	Description
query	str	—	Search terms
rows	int	10	Results per page (1-50)
offset	int	0	Pagination offset
year_from	int|None	None	Published after this year
year_to	int|None	None	Published before this year
type	str|None	None	Work type (journal-article, book-chapter)
sort	str	"relevance"	Sort: "relevance", "cited", "published"

Example prompt: "Search for papers about transformer architecture published after 2020"

---

get_paper

Get full details of a paper by DOI.

Parameter	Type	Description
doi	str	DOI identifier (e.g., 10.1038/s41586-021-03819-2)

Example prompt: "Get the details of the paper with DOI 10.1145/3295222"

Returns: title, authors, abstract, citation count, citations per year, publisher, subjects.

---

get_citations

Find papers that cite a given work.

Parameter	Type	Default	Description
doi	str	—	DOI of the cited paper
rows	int	10	Results per page (1-50)
offset	int	0	Pagination offset
sort	str	"cited"	Sort: "cited" or "published"

Example prompt: "Show me papers that cite the Attention Is All You Need paper"

---

get_author_works

Get all works by an author name.

Parameter	Type	Default	Description
author	str	—	Author name (e.g., "Yoshua Bengio")
rows	int	10	Results per page (1-50)
offset	int	0	Pagination offset
sort	str	"cited"	Sort: "cited" or "published"

Example prompt: "Find the most cited papers by Geoffrey Hinton"

Returns: publications + computed total_citations and most_cited_work.

---

get_references

Get the bibliography of a paper.

Parameter	Type	Description
doi	str	DOI of the paper

Example prompt: "What papers does this paper reference? DOI: 10.1145/3295222"

Returns: up to 30 references with DOI-resolved titles and authors where available.

---

search_by_topic

Find influential papers in a subject area.

Parameter	Type	Default	Description
topic	str	—	Subject area (e.g., "biochemistry")
rows	int	10	Results per page (1-50)
sort	str	"cited"	Sort: "cited" or "published"

Example prompt: "What are the most influential papers in reinforcement learning?"

Tech Stack

• Python 3.10+
• MCP SDK (FastMCP) — Model Context Protocol server framework
• httpx — async HTTP client with connection pooling
• Pydantic v2 — data validation and serialization
• tenacity — retry with exponential backoff
• Crossref API — scholarly metadata for 150M+ works, no auth required

Limitations

• Author search is name-based, not ORCID — may return works by different people with the same name
• Not all papers have abstracts in Crossref metadata
• Reference resolution is limited to 30 references per paper to avoid excessive API calls
• Citation counts reflect Crossref data only (may differ from Google Scholar or Semantic Scholar)
• Cache is in-memory only, cleared on server restart (30 min TTL for searches, 1 hour for individual works)

License

MIT

Author

arterorx