arXiv MCP Server

MCP (Model Context Protocol) server for searching, downloading, and analyzing arXiv papers. Supports keyword search, Semantic Scholar-backed semantic search, PDF text extraction, and SQLite caching.

Features

• Keyword Search — full arXiv API query syntax (ti:, au:, cat:, all:)
• Semantic Search — server-side semantic search via the Semantic Scholar API
• Hybrid Search — arXiv keyword results enriched with Semantic Scholar citation counts
• PDF Download — download and extract full text via PyMuPDF
• Smart Caching — SQLite cache with configurable TTL (24h metadata, 30-day PDFs)
• Categories — built-in taxonomy of ~150 arXiv categories
• Graceful Degradation — semantic/hybrid fall back to keyword when Semantic Scholar is unavailable
• Lightweight — no local ML model; depends only on mcp and pymupdf

Install

pip install arxiv-mcp-server

No model download — the server starts instantly.

Claude Code Integration

Add to ~/.claude/settings.json or .claude/settings.local.json:

{
  "mcpServers": {
    "arxiv": {
      "command": "arxiv-mcp-server"
    }
  }
}

Or from a virtual environment:

{
  "mcpServers": {
    "arxiv": {
      "command": "python",
      "args": ["-m", "arxiv_mcp_server.server"]
    }
  }
}

Configuration

All via environment variables:

Tools

search_papers

Search arXiv papers.

query: string (required) — search query
max_results: int (default 10, max 50)
start: int (default 0) — pagination offset
search_mode: "keyword" | "semantic" | "hybrid" (default "hybrid")

get_paper

Get a single paper by arXiv ID.

arxiv_id: string (required) — e.g. "1706.03762" or "1706.03762v7"

get_recent

Get recent papers in a category.

category: string (default "cs.AI")
max_results: int (default 20)

list_categories

List arXiv categories.

parent: string (optional) — filter by parent, e.g. "cs"

download_pdf

Download a paper PDF and extract text.

arxiv_id: string (required)
output_dir: string (optional) — defaults to ARXIV_CACHE_DIR/pdfs

get_cached_papers

List cached papers (diagnostic). No arguments.

Development

git clone <repo-url>
cd arxiv-mcp-server
pip install -e ".[dev]"
pytest

License

MIT