Papers MCP - Academic Papers for Any Prompt

❌ Without Papers MCP

LLMs struggle with academic papers. You get:

• ❌ Unable to effectively search across 2M+ arXiv papers
• ❌ Hallucinated paper titles, authors, and incomplete metadata
• ❌ Garbled math equations and broken tables from raw PDF parsing
• ❌ Wasted tokens from crawling full PDFs when you only need specific sections

✅ With Papers MCP

Papers MCP searches arXiv and extracts paper content as clean Markdown — directly into your LLM's context.

What are the key contributions of the "Attention Is All You Need" paper?

Find recent papers on diffusion models for image generation and summarize their methods.

• ✅ Search arXiv with full query syntax (title, author, abstract, category)
• ✅ Clean Markdown with properly formatted math equations, figures, and tables
• ✅ Extract only what you need: abstract, body, appendix, or full paper
• ✅ No more hallucinated references or broken formatting

Installation

Claude Code

claude mcp add papers-mcp -- npx papers-mcp

Claude Desktop

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

{
  "mcpServers": {
    "papers": {
      "command": "npx",
      "args": ["papers-mcp"]
    }
  }
}

Cursor

Add to your Cursor MCP config (~/.cursor/mcp.json):

{
  "mcpServers": {
    "papers": {
      "command": "npx",
      "args": ["papers-mcp"]
    }
  }
}

Available Tools

Papers MCP provides two tools following a two-step retrieval pattern:

resolve-paper-id

Resolves a paper title, author name, or search query to arXiv IDs.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | query | string | Yes | Paper title, author name, or arXiv search query |

Supports arXiv query syntax (Lucene-style field prefixes):

• all:transformer attention - Search all fields
• ti:"attention is all you need" - Search by title
• au:vaswani - Search by author
• abs:"large language model" - Search in abstract
• cat:cs.CL - Search by category

extract-paper

Retrieves detailed content from a paper using its arXiv ID.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | arxivId | string | Yes | arXiv ID (e.g., arxiv:1706.03762 or 1706.03762) | | level | enum | No | Level of detail to extract (default: body) |

Extraction levels:

• abstract - Title, authors, and abstract only
• body - Abstract + main body sections (default)
• appendix - Appendix sections only
• all - Full paper including references and appendix

Features

• Clean Markdown output - Properly formatted with headers, lists, and code blocks
• Math equation support - LaTeX equations converted to $inline$ and $$block$$ format
• Figures and tables - Preserved with captions and proper formatting
• Citations - Inline citations linked to references
• Fallback support - Tries ar5iv first, falls back to arxiv.org/html

Example Usage

Ask your LLM:

Find the original GPT paper and explain the architecture.

The LLM will:

1. Call resolve-paper-id with query "GPT language model"
2. Get back matching papers with arXiv IDs
3. Call extract-paper with the relevant arXiv ID
4. Receive the full paper content in Markdown
5. Answer your question with accurate information

License

MIT