Unpaywall MCP Server

An MCP (Model Context Protocol) server exposing Unpaywall tools so AI clients can:

• Fetch metadata by DOI
• Search article titles
• Retrieve best OA fulltext links
• Download and extract text from OA PDFs

Quickstart (npx)

Add this to your MCP client config (Claude Desktop example):

{
  "mcpServers": {
    "unpaywall": {
      "command": "npx",
      "args": ["-y", "unpaywall-mcp"],
      "env": { "UNPAYWALL_EMAIL": "[email protected]" }
    }
  }
}

Then try the tools: unpaywall_search_titles, unpaywall_get_fulltext_links, unpaywall_fetch_pdf_text.

You don't need to clone this repo or run npm install — npx handles fetching and caching on first call.

Requirements

• Node.js 18+ (for npx)
• An email address for Unpaywall / OpenAlex requests (required by Unpaywall, used for the OpenAlex polite pool).

Local development (contributors only)

End users should use the npx config above. Contributors building from source:

npm install
npm run build
[email protected] npm start   # stdio transport, as required by MCP clients

Hot-run (no build step):

[email protected] npm run dev

Tools

unpaywall_get_by_doi

• Description: Fetch Unpaywall metadata for a DOI
• Input schema:
  • doi (string, required): e.g. 10.1038/nphys1170
  • email (string, optional): overrides UNPAYWALL_EMAIL if provided
• Output: JSON response from Unpaywall

unpaywall_search_titles

• Description: Search article titles and return Unpaywall-style OA metadata for each hit (50 results/page)
• Input schema:
  • query (string, required): title query
  • is_oa (boolean, optional): if true, only OA results; if false, only closed; omit for all
  • page (integer >= 1, optional): page number
  • email (string, optional): overrides UNPAYWALL_EMAIL
• Output: JSON matching the Unpaywall search shape — results[].response is a DOI-style record (doi, title, is_oa, oa_status, best_oa_location, oa_locations), with score and snippet per result. _source: "openalex" marks the upstream.
• Note: Backed by OpenAlex's /works endpoint because Unpaywall's own /v2/search has been returning HTTP 500 since its May 2025 rewrite. Unpaywall now runs as a subroutine of OpenAlex, so this is the canonical modern equivalent.

unpaywall_get_fulltext_links

• Description: Return the best OA PDF URL and Open URL for a DOI, plus all OA locations
• Input schema:
  • doi (string, required)
  • email (string, optional): overrides UNPAYWALL_EMAIL
• Output: JSON with fields: best_pdf_url, best_open_url, best_oa_location, oa_locations, and select metadata

unpaywall_fetch_pdf_text

• Description: Download and extract text from the best OA PDF for a DOI, or from a provided pdf_url
• Input schema:
  • pdf_url (string, optional): direct PDF URL (takes precedence)
  • doi (string, optional): used to resolve best OA PDF if pdf_url not provided
  • email (string, optional): required if using doi and no UNPAYWALL_EMAIL env var
  • truncate_chars (integer >= 1000, optional): max characters of extracted text to return (default 20000)
• Output: JSON with text (possibly truncated), length_chars, truncated, pdf_url, and PDF metadata

LLM prompting tips (MCP)

When using this server from an MCP-enabled LLM client, ask the model to:

• Search then fetch: Use unpaywall_search_titles with a concise title phrase; select a result; then call unpaywall_get_fulltext_links or unpaywall_fetch_pdf_text on the chosen DOI.
• Prefer OA: Pass is_oa: true in searches when you only want open-access.
• Control size: Set truncate_chars in unpaywall_fetch_pdf_text (default 20000) and summarize long texts before proceeding.
• Be resilient: If the best PDF URL is missing, fall back to best_open_url and extract content from the landing page (outside this server).
• Respect rate limits: Space requests if making many calls; reuse earlier responses instead of repeating the same call.

Good user instructions to the LLM:

• "Find 3 OA papers about 'foundation models in biomedicine', then extract and summarize the introduction of the best one."
• "Search for 'Graph Neural Networks survey 2024', filter to OA if possible, then fetch the PDF text and produce a 10-bullet summary."

Example tool call payloads

Depending on your MCP client, the structure differs; the core payloads are:

// Search titles
{
  "name": "unpaywall_search_titles",
  "arguments": {
    "query": "graph neural networks survey",
    "is_oa": true,
    "page": 1
  }
}

// Get best OA links for a DOI
{
  "name": "unpaywall_get_fulltext_links",
  "arguments": {
    "doi": "10.48550/arXiv.1812.08434"
  }
}

// Fetch and extract PDF text (by DOI)
{
  "name": "unpaywall_fetch_pdf_text",
  "arguments": {
    "doi": "10.48550/arXiv.1812.08434",
    "truncate_chars": 20000
  }
}

Configure in an MCP client

Recommended (no-build) config for Claude Desktop using npm/npx:

{
  "mcpServers": {
    "unpaywall": {
      "command": "npx",
      "args": ["-y", "unpaywall-mcp"],
      "env": {
        "UNPAYWALL_EMAIL": "[email protected]"
      }
    }
  }
}

Alternative (local repo) config using the compiled dist:

{
  "mcpServers": {
    "unpaywall": {
      "command": "node",
      "args": ["/absolute/path/to/dist/index.js"],
      "env": {
        "UNPAYWALL_EMAIL": "[email protected]"
      }
    }
  }
}

After adding, ask your client to list tools and try:

• unpaywall_search_titles with a query
• unpaywall_get_fulltext_links with a doi
• unpaywall_fetch_pdf_text with a doi (or pdf_url)

Notes

• Respect Unpaywall's rate limits and usage guidelines: https://unpaywall.org/products/api
• The server uses stdio transport and @modelcontextprotocol/sdk.
• Set UNPAYWALL_EMAIL or pass email per call so Unpaywall can contact you about usage.

Maintainers: publish to npm

# 1) Build the project (also runs automatically on publish)
npm run build

# 2) Bump version (choose patch/minor/major)
npm version patch

# 3) Publish (ensure you are logged in: npm login)
npm publish --access public

# 4) Tag a release on GitHub (optional, recommended)

Users can then configure their MCP client with npx -y unpaywall-mcp as shown above. No clone or build required.