paper-search-mcp

paper-search-mcp is an MCP server for agents that need to search papers, read arXiv PDFs, align records across sources, and produce structured literature-analysis inputs.

The server currently integrates two paper sources:

• Semantic Scholar for citation-aware discovery and metadata lookup

• arXiv for recent papers, metadata lookup, and PDF text extraction

It also includes higher-level utilities for cross-source alignment, BibTeX export, and compact literature digests.

Chinese documentation is available in README-zh.md.

MCP Tools

search_semantic_scholar

Search Semantic Scholar and return normalized paper metadata sorted by citation count.

Parameters:

• query: Search query

• max_results: Maximum number of results, default 10

get_semantic_scholar_paper

Fetch detailed metadata for a Semantic Scholar paper by paper_id.

search_arxiv

Search arXiv and return normalized metadata.

Parameters:

• query: Search query

• max_results: Maximum number of results, default 10

• sort_by: relevance, lastUpdatedDate, or submittedDate

• sort_order: ascending or descending

get_arxiv_paper

Fetch metadata for one arXiv paper using an arXiv ID, abstract URL, or PDF URL.

read_arxiv_paper

Download an arXiv PDF, cache it locally, extract text from the first pages, and return a structured reading pack.

Parameters:

• arxiv_id_or_url: arXiv ID, abstract URL, or PDF URL

• max_pages: Maximum number of pages to extract, default 8

• max_characters: Maximum number of extracted characters, default 20000

export_bibtex

Export a paper as BibTeX.

Parameters:

• source: semantic_scholar or arxiv

• identifier: Semantic Scholar paper_id or arXiv ID/URL

align_paper_by_title

Search Semantic Scholar and arXiv by title and return exact normalized title matches across both sources.

Parameters:

• title: Paper title used for exact title alignment

• semantic_scholar_max_results: Search limit for Semantic Scholar, default 10

• arxiv_max_results: Search limit for arXiv, default 10

build_literature_digest

Search across Semantic Scholar and arXiv, deduplicate overlapping papers, and return a compact literature-analysis bundle.

This is useful for downstream agent tasks such as:

• finding classic work versus recent work

• grouping methods into families

• comparing datasets, metrics, and limitations

Related MCP server: Research Paper Ingestion MCP Server

Installation

This project is designed to use uv for environment and dependency management.

uv sync

This creates .venv in the project directory and installs the project dependencies.

To include development dependencies as well:

uv sync --group dev

If you have a Semantic Scholar API key:

export S2_API_KEY=your_key_here

Optional environment variables:

• S2_API_KEY: Semantic Scholar API key

• PAPER_MCP_HTTP_TIMEOUT: HTTP timeout in seconds, default 30

• PAPER_MCP_USER_AGENT: Custom user agent string

• PAPER_MCP_CACHE_DIR: Override the on-disk cache directory for downloaded PDFs

Install As A Python Package

For local development or direct Python-based deployment:

pip install .

To install directly from a Git repository:

pip install https://github.com/xiaoxiaoxiaotao/paper-search-mcp.git

Running The Server

Start the server directly:

uv run paper-search-mcp

Example MCP client configuration:

{
	"servers": {
		"paper-search": {
			"type": "stdio",
			"command": "uv",
			"args": [
				"run",
				"--no-sync",
				"paper-search-mcp"
			],
			"cwd": "/home/tao/code/projects/paper-search-mcp",
			"env": {
				"S2_API_KEY": "${env:S2_API_KEY}"
			}
		}
	},
	"inputs": []
}

If you prefer not to use input prompts, configure envFile instead of env and place S2_API_KEY=your_key_here in that file.

Notes

• Semantic Scholar is better for established, citation-rich papers.

• arXiv is better for recent work and full-text PDF reading.

• build_literature_digest reduces prompt assembly work for downstream agents.

• read_arxiv_paper returns text and analysis prompts instead of hard-coded conclusions.

• PDF downloads are cached on disk to avoid repeated arXiv fetches.

• An npm package is possible as a thin wrapper, but the primary runtime is still Python or Docker.

Possible Extensions

• DOI / PMID / ACL Anthology / OpenAlex support

• citation graph and related-paper retrieval

• richer section-aware PDF chunking

• persistent metadata caching beyond PDFs