paper-mcp

Remotely-callable MCP server for academic paper search, full-text retrieval & image→LaTeX, served at https://latex-tools.online/mcp.

Three corpora behind one normalized interface:

• arxiv (default) — search, metadata, and full-text (HTML / markdown / LaTeX source)

• semanticscholar (alias s2) — the full S2 API surface: citation graph, authors, recommendations, full-text snippets, bulk datasets

• openalex (alias oa) — 316M all-field works: citation graph, authors with h-index, institutions, topics, influence metrics

Plus a unified search_all that fuses all three corpora, image→LaTeX OCR, and LaTeX lint + PDF→text tooling.

Tools (41)

Generic / source-agnostic (8)

read_paper fetch chain: arxiv.org/html/{id} → ar5iv fallback (markdown/html), or arxiv.org/e-print/{id} tarball main .tex (latex). Formulas are recovered from the LaTeXML alttext invariant.

Medical / evidence-graded (1)

Image → LaTeX (3)

Turn a formula or table image back into LaTeX (e.g. a figure cropped from a paper) without needing your own vision model. Backed by the co-located recognize service (PaddleOCR-VL / DeepSeek-OCR / texify).

LaTeX tooling (3)

Companions to the LaTeX/PDF web tools at latex-tools.online — same backends, exposed over MCP.

OpenAlex (8)

• Works: get_openalex_work · get_openalex_citations · get_openalex_references · search_openalex_works (filters: year range, open-access, min-citations, institution)

• Authors/Institutions: search_openalex_authors · search_openalex_institutions

• Analytics: get_openalex_trends · list_openalex_topics

Semantic Scholar (18)

• Graph: get_paper_citations · get_paper_references · get_paper_authors

• Lookup: match_paper_title · autocomplete_papers

• Bulk: search_papers_bulk (≤1000, sortable, token paging) · get_papers_batch

• Authors: search_authors · get_author · get_author_papers · get_authors_batch

• Full-text: search_snippets (search inside paper body)

• Recommend: recommend_papers_for_paper · recommend_papers_from_examples

• Datasets: list_dataset_releases · get_dataset_release · get_dataset_download_links · get_dataset_diffs

Layout

paper_mcp/
  server.py            FastMCP server (tool registrations + instructions)
  models.py            normalized Paper model
  aggregate.py         cross-source fusion (dedup + Reciprocal Rank Fusion)
  sources/
    base.py            source registry (get_source / list_sources)
    arxiv.py           arXiv Atom API + read_paper (HTML/markdown/latex)
    semanticscholar.py Semantic Scholar full API surface
    openalex.py        OpenAlex REST API (works/authors/institutions/topics)
    recognize.py       image→LaTeX client over the co-located recognize service
    latextools.py      lint + PDF-extract clients over the latex-tools services
pyproject.toml

Run locally

cd paper-mcp
python -m venv .venv && . .venv/bin/activate
pip install -e .
PAPER_MCP_PORT=9400 python -m paper_mcp.server
# MCP endpoint at http://127.0.0.1:9400/mcp (JSON-RPC; a plain GET returns 406)

Env

Deployment (latex-tools.online)

• Runs as paper-mcp.service on tencent-us (43.130.32.180), WorkingDirectory /opt/paper-mcp, loopback port 9400.

• nginx reverse-proxies https://latex-tools.online/mcp → 127.0.0.1:9400/mcp.

• Worker-aware buckets activate only when a trusted reverse proxy overwrites X-MCP-Worker after validating the upstream platform. Never pass through a client-supplied value.

• uvicorn access logging is disabled because legacy MCP clients may put connection keys and profiles in the endpoint URL. The reverse proxy must also log $uri, not $request, for the MCP route.

• Secrets in /etc/paper-mcp.env (SEMANTIC_SCHOLAR_API_KEY).

• Runtime systemd/nginx/env files are managed by the tencent-us operations backup, not by this source repository; never commit /etc/paper-mcp.env.

Update flow

This repo is the source of truth. The server runs an independent copy under /opt/paper-mcp (not auto-synced):

# edit here → push → deploy the complete canonical Python package
rsync -a --delete paper_mcp/ tencent-us:/opt/paper-mcp/paper_mcp/
ssh tencent-us 'systemctl restart paper-mcp'
ssh tencent-us 'curl -s -o /dev/null -w "%{http_code}\n" http://127.0.0.1:9400/mcp'  # 406 = healthy (needs JSON-RPC handshake)

Production parity verified on 2026-07-23: main@07f6bbe8622aa063f56ee222a40d19c5d4264048 matches all 12 deployed Python source files byte-for-byte. The older copy embedded in latex-tools-deploy/paper-mcp/ is not a deployment source.

Notes

• arXiv calls are politely rate-limited + retried (_USER_AGENT, backoff).

• read_paper covers ~80%+ of papers via official HTML; older scan-only papers may have no full text.

• Moved here from the docs repo on 2026-06-07; that copy is gone.

License

MIT © MCPServings. See LICENSE.