research-mcp

An MCP server for academic literature research. It combines a Scopus search with three free, open scholarly APIs so you get abstracts, author metrics and open-access PDFs without needing institutional Scopus entitlements.

What it does

Citation traversal in both directions

get_citing_papers (forward) and get_references (backward) let you walk a citation graph from any starting paper. Both directions rely on OpenAlex, because a free Scopus key can serve neither:

• backward — view=REF returns 401 AUTHORIZATION_ERROR

• forward — REFEID() is a restricted search field and returns 400 INVALID_INPUT

OpenAlex covers both for free, with no quota and generally wider recall, so it is the default for each. Pass a DOI and citation traversal costs no Scopus calls at all.

Screening papers for relevance

assess_relevance takes a list of papers and a plain-language research context, and returns the evidence needed to judge each one: the abstract (from OpenAlex, which covers far more papers than CrossRef), scored topics and keywords, and which of your context's terms appear or are missing.

It screens up to 25 papers in a single call, so a whole search_scopus result set can be triaged in one turn instead of one lookup per paper.

It deliberately returns no verdict and no relevance score. The server has no language model; the judgement is left to the model calling the tool, which reads the abstract itself. The lexical_overlap field is a raw word-match signal that ignores synonyms and meaning — useful supporting evidence, not a ranking.

Why the extra APIs?

A free Scopus API key only grants the search endpoint. The Abstract Retrieval (view=FULL) and Author Retrieval endpoints require an institutional subscription and otherwise return 401 AUTHORIZATION_ERROR. To stay useful for everyone, this server falls back to:

• CrossRef for abstract text (free, no key)

• OpenAlex for author metrics incl. h-index (free, no key)

• Unpaywall for legal open-access PDFs (free, no key)

If you do have institutional Scopus access, the Scopus calls simply return richer data.

Related MCP server: Academic Paper MCP HTTP/SSE Server

Requirements

• Python ≥ 3.10

• A free Scopus API key — register at dev.elsevier.com

• Dependencies: only mcp and httpx

Getting an Elsevier (Scopus) API key

The Scopus tools need an Elsevier Developer API key. It's free to register.

Step by step

1. Create an Elsevier account. Go to dev.elsevier.com and click "I want an API Key". Sign in, or register a free account (the same account works across Elsevier products such as ScienceDirect and Mendeley).

2. Open the API Key management page. Once signed in, go to dev.elsevier.com/apikey/manage.

3. Create a new key. Click "Create API Key" and fill in:

   • Label — any name, e.g. research-mcp.

   • Website URL — required by the form but not validated for server use. You can enter http://localhost or your GitHub repo URL.

   • Optionally a CORS domain — leave blank for server-side use.

4. Accept the agreements. Tick the API Service Agreement and the text & data mining policy, then submit.

5. Copy the key. A 32-character hexadecimal key appears (e.g. 1a2b3c4d5e6f...). This is your SCOPUS_API_KEY. You can return to the manage page anytime to view or revoke it.

What the free key can and cannot do

Unlocking full Scopus data (optional)

The richer Scopus endpoints are gated by your institution's subscription, not by the key tier. To use them:

• Register / use the key from your institution's network. Elsevier binds entitlements to the institutional IP range. Create or first-use the key while on campus Wi-Fi or your university VPN.

• Alternatively, request an Institutional Token from your library and set it via the X-ELS-Insttoken header (not currently wired into this server — open an issue if you need it).

Without institutional access the server still works fully: abstracts come from CrossRef, author metrics from OpenAlex, and PDFs from Unpaywall.

Quota & fair use

• Watch your weekly quota — every call returns X-RateLimit-Remaining headers.

• Don't hammer the API in tight loops; the search tools cap at 25 results per call by design.

• The key is a secret. Keep it in the env block of your MCP config, never commit it to git (this repo's .gitignore already excludes .env and keys).

Installation

Option A — Claude Code / Claude Desktop (recommended, no server needed)

This is the simplest path. The server runs locally as a stdio process — uvx fetches it straight from GitHub, so there is no web server and no tunnel involved.

Add this to your MCP config (~/.claude.json, or the Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "research": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/niol-zh/research-mcp", "research-mcp"],
      "env": {
        "SCOPUS_API_KEY": "your-scopus-api-key",
        "UNPAYWALL_EMAIL": "you@example.com"
      }
    }
  }
}

Restart Claude. That's it — uvx installs mcp + httpx in an isolated environment automatically.

UNPAYWALL_EMAIL is required for the get_pdf_link tool. Unpaywall rejects requests without a real contact email (HTTP 422). Use your own address. All other tools work without it.

Local clone instead of GitHub

If you've cloned the repo and want to run from disk:

{
  "mcpServers": {
    "research": {
      "command": "uvx",
      "args": ["--from", "/path/to/research-mcp", "research-mcp"],
      "env": { "SCOPUS_API_KEY": "your-scopus-api-key" }
    }
  }
}

Option B — Claude Cowork (remote connector, needs an HTTPS URL)

Cowork runs in an isolated cloud VM and cannot reach a local stdio process. It needs the server exposed over HTTPS using the Streamable HTTP transport. The easiest way is a free Cloudflare Quick Tunnel.

You need two pieces running on your machine:

1. mcp-proxy — wraps the stdio server as a local HTTP server

2. cloudflared — exposes that local port over a public HTTPS URL

1. Install the helpers

# mcp-proxy comes via uvx, no install needed
# cloudflared:
#   Windows:  download cloudflared.exe from
#             https://github.com/cloudflare/cloudflared/releases/latest
#   macOS:    brew install cloudflared
#   Linux:    see Cloudflare docs

2. Start the HTTP server (terminal 1)

export SCOPUS_API_KEY="your-scopus-api-key"
export UNPAYWALL_EMAIL="you@example.com"

uvx mcp-proxy --port 8000 --transport streamablehttp \
  -e SCOPUS_API_KEY "$SCOPUS_API_KEY" \
  -- uvx --from git+https://github.com/niol-zh/research-mcp research-mcp

3. Open the tunnel (terminal 2)

cloudflared tunnel --url http://localhost:8000

Cloudflare prints a URL like https://random-words.trycloudflare.com.

4. Add the connector in Cowork

• Customize → Connectors → Add custom connector

• Name: Research

• Remote MCP server URL: https://random-words.trycloudflare.com/mcp (note the /mcp suffix)

⚠️ A Quick Tunnel URL is temporary — it changes every time you restart cloudflared, and both terminals must stay open while you use Cowork. For a stable URL, deploy the server to a host (Railway, Render, Fly.io, a VPS) or use a named Cloudflare Tunnel.

Windows users: the repo includes scripts/start-tunnel.ps1, which starts the server and the tunnel and copies the connector URL to your clipboard (requires uvx and cloudflared on PATH):

$env:SCOPUS_API_KEY = "your-key"
$env:UNPAYWALL_EMAIL = "you@example.com"
./scripts/start-tunnel.ps1

Example

A typical research flow in Claude, chaining several tools in one turn:

You: How influential is Amy Edmondson's work, and can you get me an open-access PDF of a recent open-science paper to read?

1. get_author_profile("Amy Edmondson") → OpenAlex returns disambiguated matches with metrics:

{
  "matches": [
    {
      "openalex_id": "A5003931866",
      "name": "Amy C. Edmondson",
      "h_index": 66,
      "citation_count": 45902,
      "paper_count": 358,
      "i10_index": 132,
      "affiliations": ["Harvard University"],
      "orcid": "0000-0003-4409-913X",
      "url": "https://openalex.org/A5003931866"
    }
  ],
  "note": "Top 3 matches from OpenAlex — verify by name/affiliation."
}

2. get_pdf_link("10.1371/journal.pone.0173664") → Unpaywall finds a gold open-access PDF:

{
  "doi": "10.1371/journal.pone.0173664",
  "oa_pdf_url": "https://journals.plos.org/plosone/article/file?id=10.1371/journal.pone.0173664&type=printable",
  "oa_status": "gold",
  "source": "publisher",
  "journal_is_oa": true
}

Claude: Amy C. Edmondson (Harvard) has an h-index of 66 across 358 papers with ~45,900 citations — exceptionally influential in organizational behavior. Here is an open-access PDF you can read: [link].

Every tool returns indented JSON, so Claude parses the fields reliably and combines results across calls. A full literature search would add search_scopus (to find papers) and get_abstract_details (Scopus metadata + abstract text from CrossRef).

Development

# install with dev dependencies
uv pip install -e ".[dev]"      # or: pip install -e ".[dev]"

# run the test suite (offline — all HTTP is mocked, no API key needed)
pytest

Tests live in tests/ and mock every external API with httpx.MockTransport, covering result parsing, count clamping, retry/backoff on HTTP 429, quota tracking and error paths.

Tool reference

search_scopus

get_abstract_details

get_author_profile

get_citing_papers

Forward citations — papers that cite the given document.

Why OpenAlex is the default: Scopus serves forward citations through the REFEID() search field, which is restricted. On a free key it returns HTTP 400 INVALID_INPUT ("Use of certain field restrictions in the search query is not allowed for this requestor"), so source="scopus" is only useful with an institutional subscription. If you request it anyway, the tool reports that restriction rather than leaking a bare HTTP error.

get_references

Backward citations — papers cited by the given document, most-cited first.

Reference lists come from OpenAlex, falling back to CrossRef when OpenAlex has none. For works with more than 50 references, ranking considers the first 50; total_references always reports the true count.

assess_relevance

Evidence for screening papers against a research context — no verdict, no score.

Returns per paper: title, DOI, year, venue, citation count, abstract, scored topics and keywords, and lexical_overlap (matched_terms, missing_terms, coverage). Identifiers that cannot be resolved are reported under unresolved rather than silently dropped.

Identifiers

get_references and assess_relevance accept a DOI (preferred — free), an OpenAlex work ID, or a Scopus ID. A Scopus ID costs one Scopus call to resolve its DOI first, so pass DOIs when you have them.

get_pdf_link

Scopus query syntax (quick reference)

Full syntax: Scopus Search Tips.

License

MIT — see LICENSE.

Acknowledgements

Built on the open scholarly infrastructure of CrossRef, OpenAlex and Unpaywall, plus the Elsevier Scopus API.