find-research-papers-mcp

Give your AI agent scientific grounding: search 250M+ scholarly works across five indexes with one query, and pull references and citations even from paywalled journals.

LLMs hallucinate citations. This server replaces guesswork with verified metadata — every hit is a real record from a real scholarly index, with a DOI, a URL, an abstract, and a retraction flag when the source knows one.

What it does

• One search, five indexes — arXiv, OpenAlex, Crossref, PubMed, and Semantic Scholar, aggregated into a single hit schema. The same query goes out everywhere; results come back unified.

• Paywalled papers, public bibliography — a Nature or IEEE paper you cannot read still has public metadata: get_paper returns its references (Crossref) and citing works (OpenAlex). DOI, abstract, and reference data are public even when full text is not.

• Verification built in — verify=true HEAD-checks the landing page and cross-checks OpenAlex's retraction flag. A paper that won't answer is reported as resolves: null (unknown), never as "dead".

• Graceful degradation — a rate-limited or failing source is skipped and reported in the response's skipped list. No key required anywhere; one source's outage never breaks a search.

• Zero required API keys — works out of the box. An optional Semantic Scholar key lifts its shared-pool rate limit.

Related MCP server: Academic Paper MCP HTTP/SSE Server

Sources

¹ Semantic Scholar's shared pool rate-limits without a key. Set FIND_RESEARCH_PAPERS_MCP_S2_API_KEY to lift it. When a source is skipped, the server says so in the response — it never crashes.

Tools

Install

Any MCP client (Claude Code, Cursor, opencode, …):

uvx find-research-papers-mcp        # or
npx -y find-research-papers-mcp

One-command installer (served from a Cloudflare worker — auto-detects uvx vs npx, finds your harness, merges into the right config, and reports anonymous install telemetry back to the worker — Claude Code, Cursor, opencode, Windsurf, VS Code):

curl -fsSL https://papers-mcp-install-telemetry.reachsuren.workers.dev/install?src=readme | bash
# or explicitly:  bash install.sh --claude   bash install.sh --opencode

Claude Code plugin (one-time registration, then install from anywhere):

/plugin marketplace add surendranb/find-research-papers-mcp
/plugin install find-research-papers@find-research-papers-mcp

Official MCP Registry — the server is listed as io.github.surendranb/find-research-papers-mcp, installable through clients that support the registry.

From source:

uv venv .venv --python 3.11
uv pip install --python .venv/bin/python -e .
.venv/bin/python -m papers_mcp          # stdio server

Optional config: export FIND_RESEARCH_PAPERS_MCP_S2_API_KEY=... for full-rate Semantic Scholar.

Quick start

# search everything at once
hits = search_papers("retrieval augmented generation", limit=3)
# -> unified hits: id, title, authors, year, doi, url, pdf_url,
#    citations_count, abstract, retracted, source

# deep-dive one paper, even paywalled
paper = get_paper("10.1038/s41586-023-06466-1",
                  include_references=True, include_citations=True)

# verify before you cite
v = get_paper(identifier, verify=True)["verification"]
# v["resolves"]: landing page answered?
# v["retracted"]: OpenAlex flags this paper as retracted?

A retracted: true hit must never be presented as evidence.

Telemetry & privacy

The server collects anonymous usage telemetry, on by default, to learn which sources and features matter.

• Sent: event names, an anonymous installation UUID, coarse environment signals (OS, Python version, agent name), and tool outcome counts (hits found, sources used, skipped reasons, retracted hits, latency).

• Never sent: search queries, paper results, file paths, emails, or URLs.

• Opt out with any of: FIND_RESEARCH_PAPERS_MCP_TELEMETRY=false, DISABLE_TELEMETRY=1, DO_NOT_TRACK=1, NO_TELEMETRY=1. The install ID lives in ~/.find_research_papers_mcp/installation_id; delete the folder to reset it.

• The first run prints a disclosure to stderr before anything is sent.

• Telemetry never blocks or slows the server: a dead endpoint just drops events.

Development

.venv/bin/python -m pytest tests/test_sources.py          # offline unit tests
.venv/bin/python -m pytest tests/e2e/test_e2e.py          # native MCP protocol
.venv/bin/python -m pytest -m live tests/e2e/             # live third-party APIs

License

Apache-2.0