MCP Integration

Give AI agents access to 500k+ arXiv papers via the Model Context Protocol. Our index includes all arXiv papers from 2025 onwards, plus highly-cited papers from earlier years. Connect Claude, Cursor, or any MCP client to search papers, fetch sections, and traverse citations.

What is MCP?

The Model Context Protocol is an open standard for connecting AI assistants to external data sources. ScienceStack provides an MCP server that exposes our paper database as tools that AI agents can call.

Quick Setup

You'll need an API key from Settings → API. Replace sk_live_your_key_here in the examples below.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "sciencestack": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://sciencestack.ai/api/mcp",
        "--header",
        "x-api-key:${SCIENCESTACK_API_KEY}"
      ],
      "env": {
        "SCIENCESTACK_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Restart Claude Desktop after saving.

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "sciencestack": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://sciencestack.ai/api/mcp",
        "--header",
        "x-api-key:${SCIENCESTACK_API_KEY}"
      ],
      "env": {
        "SCIENCESTACK_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Claude Code

claude mcp add sciencestack \
  --transport http \
  --url https://sciencestack.ai/api/mcp \
  --header "x-api-key: sk_live_your_key_here"

HTTP (Direct)

curl -X POST https://sciencestack.ai/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "x-api-key: sk_live_your_key_here" \
  -d '{
    "jsonrpc": "2.0",
    "method": "tools/list",
    "id": 1
  }'

Available Tools

Tool	Description	Batch Limit
`searchPapers`	Search papers by keywords	15 queries
`getPaperOverview`	Get paper structure and metadata	25 papers
`getNodes`	Fetch specific sections/equations	10 papers
`getReferences`	Get bibliography	100 refs
`getCitations`	Get forward citations	100 citations
`getAuthorPapers`	Get papers by author	100 papers
`getArtifacts`	List curated research digests	100 artifacts
`getArtifact`	Get full digest content	1 artifact
`uploadPaper`	Upload your own TeX file for parsing	1 file
`getJobStatus`	Poll upload processing status	1 job

CLI Alternative

For scripting workflows and token-efficient pipelines, consider the ScienceStack CLI:

pip install sciencestack

sciencestack search "attention mechanism" | jq '.data.papers[].title'
sciencestack nodes 1706.03762 --type equation | jq '.data.nodes[]'

The CLI outputs structured JSON envelopes, making it ideal for jq pipelines and shell scripts. MCP is better for interactive AI agents; CLI is better for automated workflows and reducing token usage.

HTML Fallback

When MCP quota is exhausted, agents can fetch individual paper elements as HTML pages:

https://sciencestack.ai/paper/1706.03762/eq/1   → Equation 1
https://sciencestack.ai/paper/1706.03762/fig/2  → Figure 2
https://sciencestack.ai/paper/1706.03762/sec/3  → Section 3

These pages render content with embedded metadata and don't require authentication. Use MCP for structured data and batch operations; use HTML fallback for quota-free access to individual elements.

Next Steps

Setup Guide — Detailed configuration for each client
Tools Reference — Full tool documentation
Examples — Real-world use cases

On this page