Research MCP

Read-only MCP server for academic research. Lets an LLM agent search papers, inspect metadata, follow citation graphs, generate BibTeX, and read full texts of open-access PDFs — all via FastMCP tools.

Requirements

• Python 3.13+
• uv
• Internet access to the public APIs (no API key required)

How it works

API	Role
Semantic Scholar	Primary search, metadata, TLDRs, citation graph. Tried first.
OpenAlex	Automatic fallback when Semantic Scholar is rate-limited, plus open-access PDF resolution.
arXiv	Dedicated preprint search (source="arxiv"), arXiv-ID lookups, and direct PDF download. Best for finding specific recent papers by exact name.
DBLP	Venue proceedings surveys (venue_proceedings) and venue-scoped searches. The canonical source for "what does this conference accept".
Crossref	BibTeX generation (transform/application/x-bibtex).
PyMuPDF	PDF text extraction.

Semantic Scholar is rate-limited when used without a key, so every request retries with exponential backoff (honoring Retry-After) before falling back to OpenAlex. Results always tell the agent which source was used. The arXiv API enforces a ~3 s spacing rule, which the server honors; DBLP resets connections on rapid sequential requests, so venue-stream pages are fetched with a 3 s politeness delay too.

Caching

All brittle API responses, search results, venue proceedings corpora, and PDFs are cached on disk under ~/.cache/research-mcp/ (api/ for JSON responses, pdf/ for downloaded PDFs):

• Fresh cache hits are served instantly — repeat queries don't touch the network.
• Venue corpora (assembled from DBLP pages) are cached as a whole, so re-surveys of a conference program are instant even if DBLP drops a page.
• Stale fallback: if an API is down or rate-limiting, the last cached response is served anyway, so the agent still gets an answer.
• Cache TTLs: 7 days for arXiv and BibTeX (stable data), 24 h for search/metadata and DBLP. Override the location with RESEARCH_MCP_CACHE_DIR.

Configuration

Register the server in your coding agent's MCP configuration. For example, in Opencode's opencode.json:

{
  "mcp": {
    "research-mcp": {
      "command": [
        "uvx",
        "--from",
        "git+https://github.com/404Simon/research-mcp",
        "research-mcp"
      ],
      "enabled": true,
      "type": "local"
    }
  }
}

Environment variables (all optional)

Variable	Purpose
SEMANTIC_SCHOLAR_API_KEY	Set for guaranteed 1 RPS and higher reliability (falls back to OpenAlex otherwise).
RESEARCH_MCP_CACHE_DIR	Where API responses and PDFs are cached. Default: ~/.cache/research-mcp/.

Tools

search_papers

Search academic papers by keyword with optional venue and year filters. Returns structured results with title, year, venue, authors, DOI, abstract, citation count, and an open-access PDF link when available.

Argument	Type	Description
query	string (req.)	Search terms
venue	string (opt.)	Filter by venue, e.g. e-Energy, VLDB
year_start	int (opt.)	Earliest publication year
year_end	int (opt.)	Latest publication year
limit	int (opt.)	Max results (default 20)
source	string (opt.)	auto (default), semanticscholar, openalex, or arxiv

source="arxiv" searches the arXiv preprint API and is the recommended way to find specific recent papers by exact name, e.g. query='ti:"carbon intensity" AND abs:forecast' (arXiv field syntax is passed through verbatim). All results are cached, so repeat searches are instant.

venue_proceedings

Survey a venue's proceedings (what actually gets published there). Use this to check whether a paper idea fits a conference like ACM e-Energy.

Argument	Type	Description
venue	string (req.)	Venue name (e.g. e-Energy) or a DBLP stream key (e.g. conf/eenergy)
year	int (opt.)	Only papers from one year (e.g. 2024)
query	string (opt.)	Topic filter; papers are ranked by title-keyword overlap
limit	int (opt.)	Max results (default 20)

The full venue corpus is fetched from DBLP (paginated past its 100-hit cap), ranked client-side, and cached as a whole — re-surveys are instant.

{"results": [
  {"title": "Reinforcement Learning Approach for Optimal Distributed Energy Management in a Microgrid",
   "year": 2018, "venue": "IEEE Transactions on Power Systems",
   "authors": ["Elham Foruzan", "Leen-Kiat Soh", "S. Asgarpoor"],
   "doi": "10.1109/tpwrs.2018.2823641",
   "abstract": "In this paper, a multiagent-based model is used ...",
   "citation_count": 297, "open_access_pdf": null,
   "source": "openalex"}],
 "note": "(Semantic Scholar unavailable — used OpenAlex instead: RequestFailed)",
 "count": 1}

paper_details

Full metadata for a paper. Accepts a DOI (e.g. 10.1109/tpwrs.2018.2823641), an arXiv ID (e.g. 2408.03506, arXiv:2408.03506, or an arxiv.org/abs/... URL), or an OpenAlex ID.

Argument	Type	Description
doi_or_id	string (req.)	DOI, arXiv ID, or OpenAlex ID

Returns title, authors, venue, year, abstract, TLDR, citation count, DOI, and open-access PDF:

get_citation_graph

Follow a paper's citations forward or backward for snowball searching.

Argument	Type	Description
paper_id	string (req.)	DOI, OpenAlex ID (W...), or Semantic Scholar paper ID
direction	string (opt.)	citing (default, papers that cite this one) or referenced (its bibliography)
limit	int (opt.)	Max results (default 20)

get_bibtex

Generate a BibTeX entry for a DOI. Uses Crossref's native transform; falls back to local generation from OpenAlex metadata (handles arXiv DOIs that Crossref doesn't know).

Argument	Type	Description
doi	string (req.)	DOI

@article{Foruzan_2018, title={Reinforcement Learning Approach for Optimal Distributed Energy Management in a Microgrid}, volume={33}, ..., author={Foruzan, Elham and Soh, Leen-Kiat and Asgarpoor, Sohrab}, year={2018} }

read_paper_full_text

Download an open-access PDF and extract its text with PyMuPDF. Pass a DOI, an arXiv ID (e.g. 2408.03506), or a direct PDF URL. PDFs are cached locally (~/.cache/research-mcp/pdf/).

Argument	Type	Description
doi_or_pdf_url	string (req.)	DOI, arXiv ID, or https://... PDF URL
max_chars	int (opt.)	Truncate returned text (default 50000)
cache	bool (opt.)	Cache the PDF locally (default true)

Returns the extracted text, page count, source URL, and cached path. If no open-access copy is found, the agent gets a clear message telling it to search for an OA copy or pass a PDF URL directly.

File structure

src/
  main.py              # FastMCP server, tool definitions, fallback orchestration
  client.py            # HTTP client with retry/backoff (429/5xx, Retry-After)
  cache.py             # Disk cache (~/.cache/research-mcp/) with stale fallback
  semanticscholar.py   # Semantic Scholar: search, details, citations (optional API key)
  openalex.py          # OpenAlex: search, details, citation graph, OA-PDF resolution
  arxiv.py             # arXiv API: search, ID lookups (3 s politeness, 7-day cache)
  dblp.py              # DBLP: venue proceedings + venue-scoped search (paged corpus cache)
  crossref.py          # Crossref: BibTeX transform + local fallback generator
  pdf.py               # PDF download/validation/caching + PyMuPDF text extraction

No API key required. Run uv sync && uv run research-mcp to start the server over stdio.