Retraction Watch API

A REST API and MCP server for searching the Retraction Watch database of retractions, corrections, expressions of concern, and reinstatements. Built with FastAPI and SQLite.

Hosted services

API

• Interactive documentation: retraction-api.onrender.com/docs
• Base URL: retraction-api.onrender.com
• Health check: retraction-api.onrender.com/health

P.S - A knowledge graph based dashboard is also present!

MCP server

• Streamable HTTP endpoint: https://retraction-watch-mcp.onrender.com/mcp
• Health check: retraction-watch-mcp.onrender.com/health

Hosted services may take a moment to wake up after being idle.

Optional research skill

The Retraction Watch research skill guides agents through citation screening and notice interpretation, including batching, source attribution, and ambiguous matches.

To install, copy the skills/retraction-watch-research/ folder from this repository into your agent's supported skills directory. Keep the folder name and SKILL.md intact, then reload skills as required by your client. Connect the MCP server below, or give the agent HTTP access to the API. Installing the skill supplies instructions; it does not configure a connection.

Try: “Use the Retraction Watch research skill to screen these citations and report matching notice types and dataset provenance.”

VS Code

Create .vscode/mcp.json:

{
  "servers": {
    "retraction-watch": {
      "type": "http",
      "url": "https://retraction-watch-mcp.onrender.com/mcp"
    }
  }
}

OpenCode

Add this to opencode.json:

{
  "mcp": {
    "retraction-watch": {
      "type": "remote",
      "url": "https://retraction-watch-mcp.onrender.com/mcp"
    }
  }
}

Claude Desktop

Open Settings → Connectors and add:

https://retraction-watch-mcp.onrender.com/mcp

What it does

• Look up papers and notices by DOI or PubMed ID, or screen up to 200 identifiers at once.
• Search titles, authors, institutions, investigation notes, and recorded misconduct reasons.
• Filter articles by journal, publisher, notice type, country, subject, and year.
• Find PubPeer links and build author or institution summaries.
• Explore journal profiles, retraction timelines, and journal/year counts with historical comparisons.
• Check the dataset's source hash, import time, and record count.

Results distinguish retractions from other notice types. A missing match does not establish that a paper is reliable. Journal/year comparisons describe recorded counts; they do not establish misconduct.

MCP tools, resources, and prompts

Tool	Result
health_check	Basic API and database connectivity
get_dataset_info	Served snapshot SHA-256, source filename, import time, row count, and importer version
list_articles	Multi-facet filtered, paginated article summaries (including has_pubpeer filter)
get_article	Full article details by record ID (including latency and pubpeer_url)
lookup_article_by_doi	Full article details by DOI (matches original paper DOI or retraction notice DOI)
lookup_article_by_pubmed	Full article details by PubMed ID (matches original paper PMID or retraction PMID)
batch_check_citations	Screen unique DOIs/PMIDs; return matching notice records, counts by notice type, and identifiers not found
search_articles	Ranked, paginated full-text search summaries
search_investigation_notes	Search full investigation narratives, institutional committee findings, and whistleblower accounts
get_pubpeer_evidence	Retrieve a PubPeer publication or search link from the record's notes
get_misconduct_taxonomy	Standardized misconduct taxonomy (image manipulation, fake peer review, paper mill, etc.)
search_by_misconduct_concept	Search articles mapped to a standardized scientific misconduct concept
search_author_retractions	Author-specific retraction records with top journals and reasons
generate_integrity_dossier	Investigative dossier for author or institution with timeline and narrative notes
analyze_retraction_timeline	Time-to-retraction delay analytics, averages, medians, and distribution brackets
detect_retraction_clusters	High-volume journal/year groups with full-group reasons and descriptive historical comparisons
get_journal_profile	Detailed journal track record, annual trend, top reasons, and latency
get_database_summary	Global database summary metrics, nature breakdown, and paywalled ratio
get_top_journals	Journals with the most retractions
get_top_reasons	Most frequently recorded retraction reasons
get_top_countries	Countries associated with the most retractions

MCP resources

Resource URI	Description
retraction://stats/summary	Global database metrics and category breakdown
retraction://stats/top-reasons	Top 25 most frequent retraction reasons
retraction://stats/top-journals	Top 25 journals with the highest retraction counts
retraction://stats/top-countries	Top 25 countries associated with retractions
retraction://stats/clusters	High-volume journal/year groups with historical comparisons
retraction://dataset	Provenance of the currently served dataset
retraction://taxonomy	Controlled scientific misconduct taxonomy and reason mappings

MCP prompts

Prompt	Description
screen_bibliography	Parses citations/DOIs, calls batch_check_citations, and generates an audit report
author_integrity_audit	Conducts a research integrity assessment on a named researcher
journal_reliability_audit	Audits a journal's track record, latency, and volume anomalies
paper_mill_investigation	Investigates high-volume clusters and coordinated paper-mill fraud
investigate_scientific_misconduct	Conducts a deep forensic investigation using misconduct taxonomy, narrative notes, and PubPeer links

Run locally

Requires Python 3.10 or later.

git clone https://github.com/ToruOkadaOi/retraction_api.git
cd retraction_api
python -m venv .venv
source .venv/bin/activate
pip install --require-hashes -r requirements-dev.lock
pip install --no-deps --no-build-isolation -e .
cp .env.example .env
python scripts/ingest_csv.py
uvicorn app.main:app --reload

Open localhost:8000/docs for the complete API reference and request examples. The importer uses the included retraction_watch.csv; run it while the API is stopped.

To run the MCP server against the local API in another terminal:

source .venv/bin/activate
RETRACTION_API_URL=http://localhost:8000 python -m mcp_server

This starts the MCP server over stdio. For HTTP transport, also set RETRACTION_MCP_TRANSPORT=streamable-http and RETRACTION_MCP_PORT=8001.

Docker

After creating .env, start both services:

docker compose up --build

API documentation: http://localhost:8000/docs · MCP endpoint: http://localhost:8001/mcp

The API image includes a validated database built from the CSV. Rebuild the image to update it.

API at a glance

Endpoint	Purpose
GET /articles	Filter and paginate article summaries
GET /articles/{record_id}	Full record details
GET /lookup/doi/{doi}	Look up an original paper or notice DOI
GET /lookup/pubmed/{pubmed_id}	Look up an original paper or notice PubMed ID
POST /lookup/batch	Screen up to 200 combined DOIs and PubMed IDs
GET /lookup/pubpeer	Find a PubPeer publication or search link
GET /search?q=...	Full-text search
GET /search/author · GET /search/dossier	Author and institution summaries
GET /search/investigation · GET /search/taxonomy · GET /search/concept/{concept}	Investigation notes and misconduct categories
GET /stats/summary · GET /stats/top-journals · GET /stats/top-reasons · GET /stats/top-countries	Database statistics
GET /stats/journal/{journal} · GET /stats/latency · GET /stats/clusters	Journal profiles, timelines, and historical comparisons
GET /dataset	Source snapshot and import details
GET /health · GET /ready	Connectivity and dataset readiness

Paginated endpoints accept skip and limit, with at most 100 records per page. Batch results contain records, counts_by_nature, and not_found_count; retracted_count counts only retraction notices.

Development

python -m ruff check app mcp_server scripts tests
python -m pytest -q

See maintenance and deployment for configuration and deployment details.