OpenPapers MCP

A local Model Context Protocol server for scientific paper research — search, metadata, abstracts, and legal Open Access PDF downloads. Runs entirely on your machine over stdio. Works with Claude Desktop, ZCode, Cursor, and any other MCP client.

Why use this instead of hitting the APIs directly? It abstracts three open scholarly APIs into five LLM-callable tools, reconstructs abstracts from OpenAlex's inverted index, finds legal OA copies via Unpaywall, and safely downloads PDFs (SSRF-safe, magic-byte-verified, atomic). No Sci-Hub, no paywall bypass, no institutional license required.

Backed by three free, open APIs

What this is not: No Sci-Hub, no paywall bypass, no EBSCO (which needs an institutional license). Only legitimately OA sources via Unpaywall.

Quick start

git clone https://github.com/Kaago/openpapers-mcp.git
cd OpenPapers
cp .env.example .env        # then edit CONTACT_EMAIL
uv sync                     # install deps
uv run pytest               # 69 offline tests (+ 3 live tests gated on OPENPAPERS_LIVE=1)
uv run openpapers           # start the stdio MCP server

Requirements: Python ≥ 3.12 and uv.

Tools

Typical workflow

search_papers("transformer attention")              → list of candidates
  └─ get_paper("10.48550/arXiv.1706.03762")         → full record + abstract
  └─ find_oa_pdf("10.48550/arXiv.1706.03762")       → OA URL
       └─ download_pdf("<url>", "10.48550/arXiv.1706.03762")  → local file

Configuration

All configuration lives in .env (git-ignored). See .env.example:

# Your email — sent only as mailto=/User-Agent to the three APIs.
# Set POLITE_POOL=0 to withhold it entirely (privacy mode).
CONTACT_EMAIL=your-email@example.com
POLITE_POOL=1

# Where downloaded PDFs land.
PDF_DIR=./pdfs

# HTTP tuning
HTTP_TIMEOUT=30
HTTP_MAX_RETRIES=3
PDF_MAX_BYTES=104857600   # 100 MB

# Logging
LOG_LEVEL=INFO

If CONTACT_EMAIL is unset, a neutral placeholder is used — the server still works, just with potentially stricter rate limits.

Privacy

• Your email is sent only in mailto= query params / User-Agent headers to OpenAlex, CrossRef, and Unpaywall — never to any other party. It is not logged at INFO level (only a masked form, e.g. you***@x.com). Set POLITE_POOL=0 to withhold it entirely.

• Search queries and DOIs are sent to those three APIs as part of normal operation. Stderr logs may include them at WARNING level on errors. The stderr stream is captured by your MCP client into a local log file.

• Downloaded PDFs persist on disk under PDF_DIR and reveal your research interests. The directory is git-ignored.

Connect your MCP client

The server is a standard stdio MCP server. Add it to your client's config, replacing /path/to/OpenPapers with your checkout path.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "openpapers": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/OpenPapers", "openpapers"]
    }
  }
}

ZCode

~/.zcode/config.toml (or workspace .zcode/config.toml):

[[mcp_servers]]
name = "openpapers"
transport = "stdio"

[mcp_servers.stdio]
command = "uv"
args = ["run", "--directory", "/path/to/OpenPapers", "openpapers"]

Cursor

.cursor/mcp.json:

{
  "mcpServers": {
    "openpapers": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/OpenPapers", "openpapers"]
    }
  }
}

After editing, restart the client. The five openpapers tools should appear.

Development

# Tests (offline — uses respx mocks, no network needed)
uv run pytest

# Lint & format
uv run ruff check .
uv run ruff format --check .

# Type check
uv run mypy

# Optional: live smoke test against the real APIs
OPENPAPERS_LIVE=1 uv run pytest -m live

See CONTRIBUTING.md for the full developer guide, including the offline-tests-only convention.

Project layout

src/openpapers/
├── __main__.py        # entry point: `uv run openpapers`
├── server.py          # FastMCP + 5 tool registrations
├── config.py          # .env loading, paths, constants, version source-of-truth
├── http_client.py     # httpx client w/ retry, polite headers, safe download
├── security.py        # SSRF & path-traversal guards
├── models.py          # Pydantic models (public tool contract)
└── services/
    ├── openalex.py    # search, get_by_doi, abstract reconstruction
    ├── crossref.py    # DOI metadata, references, funders, publisher
    ├── unpaywall.py   # OA lookup
    ├── downloader.py  # PDF download with atomic writes & magic-byte check
    └── util.py        # DOI normalization, abstract rebuild, name parsing
tests/                 # 69 offline tests via respx + 3 live smoke tests

Design notes

• Polite by default. mailto/User-Agent headers are set on every request, with a POLITE_POOL=0 escape hatch for privacy mode.

• Abstract reconstruction. OpenAlex stores abstracts as {word: [positions]} to avoid redistributing full text; the original is rebuilt losslessly.

• PDF safety. Content-type must be application/pdf and bytes must start with %PDF-. URLs are validated for scheme and resolved host (loopback, RFC1918, link-local 169.254/16, CGNAT, ULA all refused). Filenames are sanitized and paths must stay inside PDF_DIR. Downloads write to <name>.pdf.part and rename atomically on success.

• DOI normalization. Accepts bare DOIs, doi:..., https://doi.org/..., or DOIs embedded in surrounding text. The regex excludes URL-significant chars (?#&[]) to prevent API-URL injection.

• CrossRef enrichment is best-effort. If CrossRef is down or has no record, get_paper still returns the OpenAlex record with empty references.

• Single source of truth for version. __version__ lives in config.py; pyproject.toml mirrors it.

Limitations & alternatives

• No paywalled full text. If a paper isn't OA, you'll get metadata + abstract but no PDF. For paywalled content, use your institution's VPN / library proxy, or an EBSCO Discovery Service integration (requires license).

• Abstract coverage. OpenAlex has abstracts for most recent papers, but older works (pre-~2000, some publishers) may have abstract: null.

• Rate limits. The three APIs are generous but not unlimited. On 429/5xx the server retries with exponential backoff, honoring Retry-After (both seconds and HTTP-date forms). For bulk work, throttle your calls or self-host OpenAlex snapshots.

Data licensing & attribution

This server is a thin client over three upstream APIs. Each has its own terms:

• OpenAlex — metadata is CC0; they request attribution. See openalex.org.

• CrossRef — metadata is CC0; the REST API has terms of use.

• Unpaywall — see their terms.

PDFs you download come from the publishers/repositories that Unpaywall identifies; each has its own license (CC-BY, paywalled-but-OA-copy, etc.). Respect the individual paper's license.

Getting help

• 🐛 Bugs & feature requests: open an issue

• 🔒 Security issues: see SECURITY.md — do not open a public issue

• 💬 Discussion: GitHub Discussions (if enabled) or the issue tracker

License

MIT © Philipp Polte. See CONTRIBUTING.md for the contribution agreement and CODE_OF_CONDUCT.md for community standards.