paperwithcode-mcp

MCP server that brings AI paper reading and code repository discovery from Hugging Face Papers into any MCP-compatible client (Claude Desktop, IDE plugins, etc.). Supports both stdio and SSE transports.

Overview

Keeping up with AI research means reading papers, finding code implementations, and tracking daily new releases. This server bridges Hugging Face Papers' rich metadata — AI summaries, GitHub star counts, full paper markdown, and daily trending lists — directly into your AI assistant's toolset. Instead of switching between browser tabs, you query papers conversationally.

What you can do:

• Paste an arXiv ID and get the corresponding GitHub repo (with star count)
• Ask for a paper's details: title, authors, abstract, AI summary, keywords
• Read a paper's full text as markdown in your conversation
• List today's trending papers on Hugging Face Papers

Quick Start

pip install git+https://github.com/GtJerry111/paperwithcode-hf-mcp.git
paperwithcode-mcp

The server starts in stdio mode, ready to connect to Claude Desktop or any MCP host. Add it to your claude_desktop_config.json (see Claude Desktop Integration) and you're done.

Tools

resolve_code_link

Resolve an arXiv ID to its GitHub repository URL.

Parameters:

Parameter	Type	Required	Description
arxiv_id	string	Yes	The arXiv paper ID (e.g. 2508.02739)

Returns: { "github_url": "https://github.com/shiyu-coder/Kronos" }

Returns null if no GitHub repository is found for the given paper.

get_paper_details

Get detailed paper metadata from Hugging Face Papers.

Parameters:

Parameter	Type	Required	Description
arxiv_id	string	Yes	The arXiv paper ID (e.g. 2508.02739)

Returns: JSON object with:

• id — arXiv ID
• title — paper title
• authors — list of author names
• publishedAt — publication date
• summary — abstract text
• upvotes — upvote count on Hugging Face
• githubRepo — linked GitHub repository URL (if any)
• githubStars — GitHub star count (if repo exists)
• ai_summary — AI-generated summary
• ai_keywords — list of AI-extracted keywords
• discussionId — Hugging Face discussion thread ID
• markdownContentUrl — URL to the full paper markdown

Returns null if the paper is not found.

read_paper

Fetch the full text of a paper as markdown.

Parameters:

Parameter	Type	Required	Description
arxiv_id	string	Yes	The arXiv paper ID (e.g. 2508.02739)

Returns: A markdown string containing the complete paper text (abstract, introduction, method, results, etc.). Returns null if the paper cannot be found or has no markdown source.

list_daily_papers

List papers featured on Hugging Face Papers for a given date.

Parameters:

Parameter	Type	Required	Description
date	string	No	Date in YYYY-MM-DD format. Defaults to today if omitted.

Returns: A list of papers, each containing:

• id — arXiv ID
• title — paper title
• authors — list of author names
• publishedAt — publication date
• summary — abstract
• upvotes — upvote count
• numComments — number of comments on Hugging Face

Deployment

pip

pip install git+https://github.com/GtJerry111/paperwithcode-hf-mcp.git

paperwithcode-mcp                                        # stdio (default)
paperwithcode-mcp --transport sse --host 0.0.0.0 --port 8787  # SSE

uv

uv tool install git+https://github.com/GtJerry111/paperwithcode-hf-mcp.git

paperwithcode-mcp                                        # stdio

# Update later
uv tool upgrade paperwithcode-mcp

Docker

docker build -t paperwithcode-mcp .
docker run -i --rm paperwithcode-mcp                     # stdio
docker run -i --rm -p 8787:8787 paperwithcode-mcp \
  --transport sse --host 0.0.0.0 --port 8787              # SSE

Claude Desktop Integration

Add to your claude_desktop_config.json:

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

If paperwithcode-mcp is not in your PATH after pip install, use the full Python module path or the uvx launcher:

{
  "mcpServers": {
    "paperwithcode": {
      "command": "uvx",
      "args": ["paperwithcode-mcp"]
    }
  }
}

Development

git clone https://github.com/GtJerry111/paperwithcode-hf-mcp.git
cd paperwithcode-hf-mcp

# pip
pip install -e ".[dev]"

# uv
uv sync --group dev

Architecture

The server has a simple data flow:

MCP tool call -> mcp_server.py (FastMCP) -> resolver.py (business logic)
    -> client.py (curl/network) + parser.py (HTML extraction)

• mcp_server.py — FastMCP instance with 4 tool definitions and the CLI entry point
• resolver.py — orchestrates calls between client and parser, returns typed results
• client.py — PaperPageClient wraps curl subprocess, handles proxy and retries
• parser.py — extracts structured data from Hugging Face paper pages

Data Sources

• https://huggingface.co/papers/{arxiv_id} — individual paper page (embedded JSON in data-props)
• https://huggingface.co/api/daily_papers?date=YYYY-MM-DD — daily papers API (no auth)
• markdownContentUrl — full paper text as markdown from the arXiv HTML conversion

Environment Variables

Variable	Default	Description
HTTPS_PROXY / HTTP_PROXY / ALL_PROXY	—	Proxy for outgoing HTTP requests
PWC_TIMEOUT	15.0	Request timeout in seconds

Limitations

This project uses Hugging Face Papers as its data source, NOT the paperswithcode.com API (which is no longer available). As a result:

• No paper search by keyword or title
• No conference, proceedings, or author browsing
• No benchmark results or dataset listings

License

MIT