refcheck

Academic reference verification for AI assistants.

An MCP server that cross-checks academic citations against real publication databases. It catches hallucinated references, finds real papers on a topic, and exports publication-ready BibTeX — all from within Claude Code, Claude Desktop, or any MCP-compatible client.

The Problem

Large language models frequently hallucinate academic citations: plausible-looking references with fabricated titles, authors, DOIs, or venues. A single hallucinated citation in a paper or grant proposal can undermine credibility.

refcheck solves this by giving your AI assistant direct access to Crossref, Semantic Scholar, and arXiv. Every reference it returns is backed by a real database record.

What It Does

Tool	Purpose
verify_reference	Check if a citation is real. Returns verified, partial_match, or not_found with a confidence score and corrected BibTeX.
search_references	Find real papers on a topic. Never fabricates — only returns database-backed results.
get_bibtex	Export publication-ready BibTeX by DOI, title, or Semantic Scholar ID. Supports batch export.

Quick Start

1. Install

pip install refcheck-mcp

Or install from source:

git clone https://github.com/benchoi93/refcheck.git
cd refcheck
pip install -e .

Requirements: Python 3.11+

2. Configure Claude Code

Add to your Claude Code settings (~/.claude.json):

{
  "mcpServers": {
    "refcheck": {
      "command": "python",
      "args": ["-m", "refcheck"],
      "env": {
        "CROSSREF_EMAIL": "you@university.edu"
      }
    }
  }
}

3. Use It

> Verify this reference: Li, Y., Yu, R., Shahabi, C., & Liu, Y. (2018).
  Diffusion convolutional recurrent neural network: Data-driven traffic
  forecasting. ICLR 2018.

> Find me 5 recent papers on physics-informed neural networks for traffic
  state estimation and get their BibTeX.

> Get BibTeX for DOI 10.1016/j.trc.2024.104873

> Verify all references in my refs.bib and replace any incorrect entries
  with the corrected BibTeX.

Tools

verify_reference

Checks whether a citation is real by querying multiple databases and fuzzy-matching the results.

Parameters:

Parameter	Type	Required	Description
title	str	At least one of title or doi	Paper title
authors	list[str]	No	Author names (last names sufficient)
year	int	No	Publication year
doi	str	At least one of title or doi	DOI string
venue	str	No	Journal or conference name

Verification pipeline:

1. DOI lookup (if provided) — Resolves via Crossref. A DOI match is near-certain verification.
2. Title search — Queries all available databases in parallel.
3. Fuzzy matching — Scores each candidate on title similarity (token set ratio), author overlap (Jaccard on last names), year match (+/-1 tolerance), and venue similarity.
4. Verdict — Weighted composite score produces a classification:

Verdict	Score	Meaning
verified	>= 0.85	High confidence the reference is real and accurate
partial_match	0.50 - 0.85	A similar paper exists but some fields differ
not_found	< 0.50	No match in any database. Likely hallucinated.

Automatic BibTeX correction: When the verdict is verified or partial_match, the response includes a corrected_bibtex field with the correct BibTeX entry from the matched paper. This means a single call both detects errors and provides the fix.

Example response (partial match):

{
  "verdict": "partial_match",
  "confidence": 0.72,
  "matched_reference": {
    "title": "Attention is All You Need",
    "authors": ["Ashish Vaswani", "Noam Shazeer", "..."],
    "year": 2017,
    "doi": "10.48550/arXiv.1706.03762",
    "venue": "Advances in Neural Information Processing Systems",
    "url": "https://doi.org/10.48550/arXiv.1706.03762"
  },
  "discrepancies": ["Year mismatch: claimed 2018, found 2017"],
  "sources_checked": ["crossref", "semantic_scholar", "arxiv"],
  "corrected_bibtex": "@article{vaswani2017attention,\n  author = {Vaswani, Ashish and ...},\n  title = {Attention is All You Need},\n  year = {2017},\n  doi = {10.48550/arXiv.1706.03762},\n}"
}

search_references

Searches for real papers across all configured databases. Results are deduplicated by DOI and title similarity.

Parameters:

Parameter	Type	Default	Description
query	str	required	Search topic or keywords
max_results	int	10	Number of results (1-30)
year_from	int	None	Earliest publication year
year_to	int	None	Latest publication year
databases	list[str]	All available	Which databases to query

Each result includes: title, authors, year, DOI, venue, abstract (truncated), URL, and source database.

get_bibtex

Generates verified BibTeX entries. Prefers Crossref content negotiation (publisher-quality BibTeX) and falls back to metadata construction when needed.

Parameters (provide one):

Parameter	Type	Description
doi	str	Single DOI
dois	list[str]	Multiple DOIs (batch mode)
title	str	Paper title (searches and verifies first)
semantic_scholar_id	str	Semantic Scholar paper ID

Example — batch DOI export:

get_bibtex(dois=["10.1016/j.trc.2024.104873", "10.48550/arXiv.1706.03762"])

Returns concatenated BibTeX entries ready to append to a .bib file:

@article{vaswani2017attention,
  author        = {Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and ...},
  title         = {Attention is All You Need},
  journal       = {Advances in Neural Information Processing Systems},
  year          = {2017},
  doi           = {10.48550/arXiv.1706.03762},
  url           = {https://doi.org/10.48550/arXiv.1706.03762},
}

BibTeX conventions:

• Citation keys follow {lastname}{year}{keyword} format (e.g., vaswani2017attention)
• Acronyms and proper nouns are wrapped in {} to preserve capitalization
• Entry types are inferred from metadata (@article, @inproceedings, @misc, etc.)
• Fields are never fabricated. Missing data (e.g., page numbers) is omitted rather than guessed.

Databases

refcheck queries multiple academic databases in parallel. It works out of the box with no API keys — Crossref and arXiv are free and unauthenticated. Adding optional keys unlocks additional sources and better rate limits.

Database	Key Required	Coverage	Notes
Crossref	No	156M+ DOI-registered publications	DOI backbone. Set CROSSREF_EMAIL for faster rate limits.
Semantic Scholar	No	200M+ papers	Broadest coverage. Free API key recommended.
arXiv	No	2.5M+ preprints	Essential for recent ML/CS work. 3s rate limit enforced.
Scopus	Yes	90M+ records	Elsevier API key + optional institutional token.
IEEE Xplore	Yes	IEEE/IET publications	Free tier: 200 requests/day.

The server operates in degraded mode when keys are missing — it skips unavailable APIs and reports which sources were checked.

Configuration

Copy the example env file and fill in your keys:

cp .env.example .env

Then edit .env:

# Strongly recommended — enables Crossref "polite pool" (~50 req/sec)
CROSSREF_EMAIL=you@university.edu

# Recommended — free key from https://www.semanticscholar.org/product/api
S2_API_KEY=your-s2-key

# Optional — Elsevier API key from https://dev.elsevier.com/
ELSEVIER_KEY=
# Optional — Elsevier institutional token (higher rate limits)
ELSEVIER_INSTTOKEN=

# Optional — free key from https://developer.ieee.org/
IEEE_API_KEY=

The .env file is gitignored. The server loads it automatically on startup via python-dotenv.

Alternatively, you can set environment variables directly or pass them in the env block of your MCP server config in ~/.claude.json.

Architecture

refcheck/
├── src/refcheck/              # Python package
│   ├── __init__.py            # Package version
│   ├── __main__.py            # python -m refcheck entry point
│   ├── server.py              # FastMCP server, three tool definitions
│   ├── clients/
│   │   ├── __init__.py        # Client protocol & registry
│   │   ├── crossref.py        # Crossref: DOI lookup, title search, BibTeX export
│   │   ├── semantic_scholar.py # Semantic Scholar: keyword search, paper lookup
│   │   ├── arxiv.py           # arXiv: title/keyword search, XML parsing
│   │   ├── scopus.py          # Scopus: search with API key (optional)
│   │   └── ieee.py            # IEEE Xplore: search with API key (optional)
│   ├── matching.py            # Fuzzy matching: title, author, year, venue scoring
│   ├── bibtex.py              # BibTeX generation, formatting, citation keys
│   ├── models.py              # Pydantic v2 input/output models
│   └── config.py              # Environment variable management
├── pyproject.toml             # Package metadata & dependencies
├── server.json                # MCP Registry metadata
└── requirements.txt           # Pinned dependencies (alternative to pyproject.toml)

Key design decisions:

• Async throughout — All API calls use httpx.AsyncClient with 10-second timeouts.
• Parallel queries — Multiple databases are queried simultaneously via asyncio.gather.
• Graceful degradation — One failing API never blocks others. Errors are logged to stderr and the response indicates which sources were checked.
• Never fabricates — If no match is found, the tool says so. BibTeX fields are omitted rather than guessed.

Use Cases

• Writing papers — Verify every citation before submission. Catch wrong years, misspelled authors, or entirely fabricated references. Partial matches return corrected BibTeX automatically.
• Fixing .bib files — Verify references in bulk and replace incorrect entries with database-backed corrected BibTeX in a single pass.
• Literature reviews — Search for real papers on a topic and get BibTeX in one step.
• Grant proposals — Ensure all referenced prior work actually exists.
• Teaching — Help students verify citations in their assignments.
• Batch BibTeX export — Collect DOIs from a draft and export a clean .bib file.

Limitations

• Coverage gaps — Some niche publications may not appear in the queried databases. A not_found result does not guarantee the paper is fake; it may exist in databases not covered (e.g., PubMed for biomedical, DBLP for CS).
• Metadata quality — BibTeX constructed from metadata (when Crossref content negotiation is unavailable) may be missing fields like page numbers or volume.
• Rate limits — arXiv enforces a 3-second delay between requests. IEEE free tier is limited to 200 requests/day.
• No Google Scholar — There is no official Google Scholar API. The existing coverage from Crossref + Semantic Scholar + arXiv is broader and more reliable.

License

MIT

Citation

If you use refcheck in your research workflow, a mention is appreciated:

refcheck: Academic Reference Verification MCP Server
https://github.com/benchoi93/refcheck