Top Paper MCP Server

🔍 Enable AI assistants to search and access academic papers from arXiv and top conferences through a simple MCP interface.

The Top Paper MCP Server provides a bridge between AI assistants and academic research repositories (arXiv, CVPR, NeurIPS, ICLR, ICML, etc.) through the Model Context Protocol (MCP). It allows AI models to search for papers and access their content in a programmatic way.

✨ Core Features

arXiv Integration

• 🔎 Paper Search: Query arXiv papers with filters for date ranges and categories
• 📄 Paper Access: Download and read paper content
• 📋 Paper Listing: View all downloaded papers
• 🗃️ Local Storage: Papers are saved locally for faster access

Conference Support

• 🔎 Conference Search: Search papers from top AI/ML/CV/NLP conferences:
  • CVF: CVPR, ICCV, WACV
  • ECVA: ECCV
  • OpenReview: ICLR, NeurIPS, ICML, AAAI, IJCAI, ACL, EMNLP, NAACL, COLM, CoRL, MLSYS, MICCAI, IWSLT, INTERSPEECH
  • ML Anthology: COLT, UAI
  • ACM: ACM Digital Library (SIGGRAPH, CHI, KDD, etc.)
• 📄 Paper Download: Download papers from conference websites
• 📝 Prompts: Research prompts for paper analysis

Supported Conferences

Conference	Data Source	Year Range
Computer Vision
CVPR	CVF Open Access	2000-present
ICCV	CVF Open Access	2000-present
WACV	CVF Open Access	2000-present
ECCV	ECVA	2000-present
Machine Learning / AI
ICLR	OpenReview API	2000-present
NeurIPS	OpenReview API	2000-present
ICML	OpenReview API	2000-present
AAAI	OpenReview API	2000-present
IJCAI	OpenReview API	2000-present
COLM	OpenReview API	2000-present
CoRL	OpenReview API	2000-present
MLSYS	OpenReview API	2020-present
NLP
ACL	OpenReview API	2000-present
EMNLP	OpenReview API	2000-present
NAACL	OpenReview API	2000-present
Speech / Multimodal
INTERSPEECH	OpenReview API	2000-present
IWSLT	OpenReview API	2000-present
MICCAI	OpenReview API	2000-present
Theory
COLT	ML Anthology	2000-present
UAI	ML Anthology	2000-present
Other
ACM	ACM Digital Library	Varies

🚀 Quick Start

Installing via Smithery

npx -y @smithery/cli install top-paper-mcp-server --client claude

Manual Installation

uv tool install top-paper-mcp-server

For PDF support (older papers):

uv tool install 'top-paper-mcp-server[pdf]'

Verify installation:

top-paper-mcp-server --help

MCP Configuration

{
    "mcpServers": {
        "top-paper": {
            "command": "uv",
            "args": [
                "tool",
                "run",
                "top-paper-mcp-server",
                "--storage-path", "/path/to/paper/storage"
            ]
        }
    }
}

For development:

{
    "mcpServers": {
        "top-paper": {
            "command": "uv",
            "args": [
                "--directory",
                "path/to/cloned/top-paper-mcp-server",
                "run",
                "top-paper-mcp-server",
                "--storage-path", "/path/to/paper/storage"
            ]
        }
    }
}

HTTP Transport

TRANSPORT=http HOST=127.0.0.1 PORT=8080 top-paper-mcp-server --storage-path /path/to/papers

Then configure your MCP client:

{
    "mcpServers": {
        "top-paper": {
            "type": "http",
            "url": "http://127.0.0.1:8080/mcp"
        }
    }
}

💡 Available Tools

arXiv Tools

# Search arXiv papers
result = await call_tool("search_papers", {
    "query": "transformer",
    "max_results": 10,
    "categories": ["cs.LG", "cs.AI"]
})

# Download a paper
result = await call_tool("download_paper", {
    "paper_id": "2401.12345"
})

# List downloaded papers
result = await call_tool("list_papers", {})

# Read paper content
result = await call_tool("read_paper", {
    "paper_id": "2401.12345"
})

Conference Tools

# Search single conference
result = await call_tool("conference_search", {
    "query": "object detection",
    "conference": "CVPR",
    "year": 2024,
    "max_results": 10
})

# Multi-conference concurrent search (NEW!)
result = await call_tool("conference_search", {
    "query": "transformer",
    "conference": "NeurIPS",
    "year": 2024,
    "search_all": True,
    "conferences": ["CVPR", "NeurIPS", "ICLR", "ICML"]
})

# Search by category with concurrent execution
result = await call_tool("conference_search", {
    "query": "attention",
    "conference": "NeurIPS",
    "year": 2024,
    "search_all": True,
    "categories": ["computer_vision", "nlp"]
})

# Unified search across ALL conferences
result = await call_tool("unified_search", {
    "query": "deep learning",
    "year": 2024,
    "max_results_per_conference": 5,
    "total_results": 20
})

# Download conference paper
result = await call_tool("conference_download", {
    "paper_id": "12345",
    "conference": "CVPR",
    "year": 2024
})

Multi-Threaded Concurrent Search Features

• Concurrent Execution: Searches multiple conferences in parallel using asyncio
• Priority-Based Ordering: Results sorted by conference priority (CVPR > NeurIPS > ICLR > ICML > ...)
• Category Filtering: Filter by domain (computer_vision, machine_learning, nlp, ai, speech, medical, theory)
• Results Aggregation: Merges and deduplicates results from multiple sources
• Rate Limiting: Built-in semaphore limits concurrent requests (max 10) to prevent API throttling
• Timeout Protection: Individual requests timeout after 30 seconds to prevent slow endpoints blocking
• Automatic Retry: Failed requests automatically retry up to 2 times with exponential backoff
• Error Resilience: Graceful handling of partial failures - successful results still returned

⚙️ Configuration

Setting	Purpose	Default
--storage-path	Paper storage location	~/.top-paper-mcp-server/papers
MAX_RESULTS	Maximum search results	50
REQUEST_TIMEOUT	API timeout in seconds	60
TRANSPORT	Transport type: stdio, http	stdio
HOST	Host to bind to in HTTP mode	127.0.0.1
PORT	Port to listen on in HTTP mode	8000

🔒 Security

Paper content retrieved from external sources is untrusted input.

When an AI assistant downloads or reads a paper through this server, the paper's text is passed directly into the model's context. A maliciously crafted paper could embed adversarial instructions designed to hijack the AI's behavior.

Recommended Mitigations

1. Use read-only MCP configurations when possible
2. Review paper content before acting on AI summaries
3. Be cautious in multi-tool setups
4. Treat AI-generated summaries as data, not instructions

🧪 Testing

python -m pytest

📄 License

Released under the Apache License 2.0. See the LICENSE file for details.

---

🙏 Acknowledgments

This project is built upon the excellent foundation of arxiv-mcp-server created by Joseph Blazick.

We sincerely thank:

• Joseph Blazick for creating the original arxiv-mcp-server and making it open source
• The arXiv team for providing the open research repository
• OpenReview for enabling open access to peer reviews
• CVF/ECVA for providing open access to computer vision conference papers
• All the conference organizers who make their papers publicly accessible

This project extends the original work to support additional top academic conferences while maintaining compatibility with the existing arXiv functionality.

---