ArXiv Research MCP Server 🔬

The MCP server every data scientist needs. Search 2M+ ArXiv papers, find code implementations, track citations, get SOTA benchmarks — all from your AI assistant.

Why This Server?

Data scientists spend ~30% of their time searching for papers, checking if code exists, comparing benchmarks. This MCP server brings all of that into your LLM workflow with zero API keys required.

Tools Available

Tool	What it does
arxiv_search_papers	Full-text search across 2M+ ArXiv papers with category filtering
arxiv_get_paper	Complete paper details with citations + code repos enrichment
arxiv_get_citations	Citation analysis via Semantic Scholar (citing papers, references)
arxiv_find_code	Find GitHub repos implementing a paper (Papers with Code)
arxiv_trending_papers	Latest papers in any ArXiv category
arxiv_get_benchmarks	SOTA benchmark lookup by task or dataset
arxiv_semantic_search	Semantic search via Semantic Scholar (better than keyword search)
arxiv_list_categories	List all supported ArXiv categories

Quick Start

Option 1: Connect directly (Streamable HTTP)

Add to your Claude Desktop / MCP client config:

{
  "mcpServers": {
    "arxiv-research": {
      "type": "url",
      "url": "http://localhost:8000/mcp"
    }
  }
}

Option 2: Run locally

pip install -r requirements.txt
python arxiv_research_mcp.py

Option 3: Docker

docker build -t arxiv-research-mcp .
docker run -p 8000:8000 arxiv-research-mcp

Option 4: Using uv (recommended)

uv run --with "mcp[cli]" --with httpx --with pydantic arxiv_research_mcp.py

Example Queries

Once connected, ask your AI assistant:

• "Search ArXiv for papers on LoRA fine-tuning from 2024"
• "Get full details for paper 2106.09685 including citations and code"
• "Find code implementations for the Mamba architecture paper"
• "What are the trending papers in cs.CL this week?"
• "Show SOTA benchmarks for object detection"
• "Semantic search for papers on retrieval augmented generation"

APIs Used (All Free, No Keys)

API	Rate Limit	Data
ArXiv API	1 req/3s	2M+ papers, full metadata
Semantic Scholar	~100 req/5min	Citations, references, TL;DR
Papers with Code	Fair use	Code repos, benchmarks, tasks

Deployment

Railway / Render / Fly.io

The server runs on port 8000 with streamable HTTP transport. Deploy the Dockerfile to any container hosting platform. No environment variables or API keys needed.

Permanent self-hosting

docker compose up -d

License

MIT — use it, fork it, make it better.

Author

Dr. Swapnil Morande — Principal AI Architect