crossref-local 0.7.4

Local CrossRef database with 167M+ works and full-text search

CrossRef Local (crossref-local)

Local CrossRef database with 167M+ scholarly works, full-text search, and impact factor calculation

Demo

# Search 167M papers locally — no API rate limits, ~22 ms full-text query
crossref-local search "epilepsy seizure prediction"

# Resolve a DOI to full record (title, abstract, citations, journal IF)
crossref-local search-by-doi 10.1038/nature11247

# Drive from MCP / Claude Code
crossref-local mcp serve

The image is a live capture against the local DB; the <details> block below has a 6m55s MCP-driven demo video.

Architecture

┌──────────────────────────┐    ┌──────────────────────────┐
│ CrossRef public dump     │    │ JCR / OpenAlex IF tables │
│ (~100 GB compressed)     │    │                          │
└──────────────┬───────────┘    └──────────────┬───────────┘
               │ dois2sqlite                   │
               ▼                               ▼
       ┌─────────────────┐               ┌──────────────┐
       │ crossref.db     │ ◀── joins ──▶ │ impact-factor│
       │ (SQLite + FTS5) │               │ table        │
       └────────┬────────┘               └──────────────┘
                │
                ▼
   ┌──────────────────────────────────┐
   │ crossref-local — Python / CLI / MCP │
   │   search · search-by-doi · cache    │
   │   stats · check-citations · relay   │
   └──────────────────────────────────┘

The DB lives entirely on disk; crossref-local is a thin facade over SQLite + FTS5 + a small impact-factor table. No network calls during queries; rebuild scripts under make fts-build-screen / citations-build-screen are the only producers of state.

MCP Demo Video

Live demonstration of MCP server integration with Claude Code for epilepsy seizure prediction literature review:

• Full-text search on title, abstracts, and keywords across 167M papers (22ms response)

📄 Full demo documentation | 📊 Generated diagrams

Why CrossRef Local?

Built for the LLM era - features that matter for AI research assistants:

Feature	Benefit
📝 Abstracts	Full text for semantic understanding
📊 Impact Factor	Filter by journal quality
🔗 Citations	Prioritize influential papers
⚡ Speed	167M records in ms, no rate limits

Perfect for: RAG systems, research assistants, literature review automation.

Installation

pip install crossref-local

From source:

git clone https://github.com/ywatanabe1989/crossref-local
cd crossref-local && make install

Database setup (1.5 TB, ~2 weeks to build):

# 1. Download CrossRef data (~100GB compressed)
aria2c "https://academictorrents.com/details/..."

# 2. Build SQLite database (~days)
pip install dois2sqlite
dois2sqlite build /path/to/crossref-data ./data/crossref.db

# 3. Build FTS5 index (~60 hours) & citations table (~days)
make fts-build-screen
make citations-build-screen

Python API

from crossref_local import search, get, count

# Full-text search (22ms for 541 matches across 167M records)
results = search("hippocampal sharp wave ripples")
for work in results:
    print(f"{work.title} ({work.year})")

# Get by DOI
work = get("10.1126/science.aax0758")
print(work.citation())

# Count matches
n = count("machine learning")  # 477,922 matches

Async API:

from crossref_local import aio

async def main():
    counts = await aio.count_many(["CRISPR", "neural network", "climate"])
    results = await aio.search("machine learning")

CLI

crossref-local search "CRISPR genome editing" -n 5
crossref-local search-by-doi 10.1038/nature12373
crossref-local status  # Configuration and database stats

With abstracts (-a flag):

$ crossref-local search "RS-1 enhances CRISPR" -n 1 -a

Found 4 matches in 128.4ms

1. RS-1 enhances CRISPR/Cas9- and TALEN-mediated knock-in efficiency (2016)
   DOI: 10.1038/ncomms10548
   Journal: Nature Communications
   Abstract: Zinc-finger nuclease, transcription activator-like effector nuclease
   and CRISPR/Cas9 are becoming major tools for genome editing...

HTTP API

Start the FastAPI server:

crossref-local relay --host 0.0.0.0 --port 31291

Endpoints:

# Search works (FTS5)
curl "http://localhost:31291/works?q=CRISPR&limit=10"

# Get by DOI
curl "http://localhost:31291/works/10.1038/nature12373"

# Batch DOI lookup
curl -X POST "http://localhost:31291/works/batch" \
  -H "Content-Type: application/json" \
  -d '{"dois": ["10.1038/nature12373", "10.1126/science.aax0758"]}'

# Citation endpoints
curl "http://localhost:31291/citations/10.1038/nature12373/citing"
curl "http://localhost:31291/citations/10.1038/nature12373/cited"
curl "http://localhost:31291/citations/10.1038/nature12373/count"

# Collection endpoints
curl "http://localhost:31291/collections"
curl -X POST "http://localhost:31291/collections" \
  -H "Content-Type: application/json" \
  -d '{"name": "my_papers", "query": "CRISPR", "limit": 100}'
curl "http://localhost:31291/collections/my_papers/download?format=bibtex"

# Database info
curl "http://localhost:31291/info"

HTTP mode (connect to running server):

# On local machine (if server is remote)
ssh -L 31291:127.0.0.1:31291 your-server

# Python client
from crossref_local import configure_http
configure_http("http://localhost:31291")

# Or via CLI
crossref-local --http search "CRISPR"

MCP Server

Run as MCP (Model Context Protocol) server:

crossref-local mcp start

Local MCP client configuration:

{
  "mcpServers": {
    "crossref-local": {
      "command": "crossref-local",
      "args": ["mcp", "start"],
      "env": {
        "CROSSREF_LOCAL_DB": "/path/to/crossref.db"
      }
    }
  }
}

Remote MCP via HTTP (recommended):

# On server: start persistent MCP server
crossref-local mcp start -t http --host 0.0.0.0 --port 8082

{
  "mcpServers": {
    "crossref-remote": {
      "url": "http://your-server:8082/mcp"
    }
  }
}

Diagnose setup:

crossref-local mcp doctor        # Check dependencies and database
crossref-local mcp list-tools    # Show available MCP tools
crossref-local mcp installation  # Show client config examples

See docs/remote-deployment.md for systemd and Docker setup.

Available tools:

• search - Full-text search across 167M+ papers
• search_by_doi - Get paper by DOI
• enrich_dois - Add citation counts and references to DOIs
• status - Database statistics
• cache_* - Paper collection management

Impact Factor

from crossref_local.impact_factor import ImpactFactorCalculator

with ImpactFactorCalculator() as calc:
    result = calc.calculate_impact_factor("Nature", target_year=2023)
    print(f"IF: {result['impact_factor']:.3f}")  # 54.067

Journal	IF 2023
Nature	54.07
Science	46.17
Cell	54.01
PLOS ONE	3.37

Citation Network

from crossref_local import get_citing, get_cited, CitationNetwork

citing = get_citing("10.1038/nature12373")  # 1539 papers
cited = get_cited("10.1038/nature12373")

# Build visualization (like Connected Papers)
network = CitationNetwork("10.1038/nature12373", depth=2)
network.save_html("citation_network.html")  # requires: pip install crossref-local[viz]

Performance

Query	Matches	Time
hippocampal sharp wave ripples	541	22ms
machine learning	477,922	113ms
CRISPR genome editing	12,170	257ms

Searching 167M records in milliseconds via FTS5.

Related Projects

openalex-local - Sister project with OpenAlex data:

Feature	crossref-local	openalex-local
Works	167M	284M
Abstracts	~21%	~45-60%
Update frequency	Real-time	Monthly
DOI authority	✓ (source)	Uses CrossRef
Citations	Raw references	Linked works
Concepts/Topics	❌	✓
Author IDs	❌	✓
Best for	DOI lookup, raw refs	Semantic search

When to use CrossRef: Real-time DOI updates, raw reference parsing, authoritative metadata. When to use OpenAlex: Semantic search, citation analysis, topic discovery.

---

Installation

Recommended: uv pip install crossref-local[all] — uv's Rust resolver handles the SciTeX dep set in 1-3 min where pip's serial backtracker can take 30+ min on the full extras. Plain pip install still works; the install block below shows both.

pip install crossref-local              # core
pip install crossref-local[mcp]         # + MCP server

4 Interfaces

Python API

from crossref_local import crossref_search, get_work

results = crossref_search("deep learning EEG", limit=10)
work = get_work("10.1038/nature12373")

CLI

crossref-local search "query"
crossref-local doi 10.1038/nature12373

MCP Server

crossref-local mcp start

Skills

Agent skill pages live under src/crossref_local/_skills/crossref-local/.

Problem and Solution

#	Problem	Solution
1	CrossRef public API is rate-limited + requires internet + slow for bulk queries -- 167M works is the bottleneck for literature tools	Local SQLite + FTS5 -- full CrossRef dump (~60 GB) queryable offline; crossref_search returns in milliseconds

Part of SciTeX

crossref-local is part of SciTeX. Install via the umbrella with pip install scitex[scholar] to use as scitex.scholar (Python) or scitex scholar ... (CLI) — crossref-local provides the local CrossRef backing for scholar's DOI resolution.

import scitex

scitex.scholar.enrich_bibtex("references.bib")
scitex.scholar.check_citations("manuscript.tex")

Four Freedoms for Research

0. The freedom to run your research anywhere — your machine, your terms.
1. The freedom to study how every step works — from raw data to final manuscript.
2. The freedom to redistribute your workflows, not just your papers.
3. The freedom to modify any module and share improvements with the community.

AGPL-3.0 — because we believe research infrastructure deserves the same freedoms as the software it runs on.

---