RefExists

https://refexists.dsl4.com — do your references actually exist?

Paste a bibliography (or any text containing citations). Your browser checks every reference directly against the world's scholarly registries — Crossref, DataCite, doi.org, OpenAlex, PubMed, Open Library — and stamps each one with an evidence-backed verdict. It catches:

• Fabricated DOIs — identifiers that no registration agency has ever issued
• AI-hallucinated references — plausible-looking citations with no matching record in 250M+ indexed works
• Stitched citations — a real DOI attached to a different paper's title/authors (the classic LLM fabrication signature)
• Retracted works cited as if they were valid evidence

Why it's different

Every comparable checker (2024–2026 wave: CiteSure, AiCitationChecker, CiteMe, GPTZero hallucination check, Paperpile's checker) uploads your text to their server, and most require accounts and cap free usage. RefExists is 100% client-side: a static page, no accounts, no analytics, no server, no caps beyond the registries' own public rate budgets. Court filings, unpublished manuscripts and misconduct-review documents never leave your machine except as individual lookups to the public registries — enforced by the page's Content-Security-Policy, not just promised.

Honesty model

• A network failure is never reported as "not found" — absence verdicts require a definitive registry 404 (Crossref → DataCite → the doi.org handle system, which is authoritative for every DOI agency).
• "No record" is explicitly framed as a lead, not a conviction (books, theses, non-English and very recent works are under-indexed).
• Web-only references are "Can't check", never accused.
• "Verified" means the reference exists and matches — the page states plainly that it does not mean the source supports the claim it's cited for.
• Every verdict shows its evidence trail (which registries said what, with links).

Run locally

python3 -m http.server 8471 --directory public
# open http://localhost:8471

No build step, no dependencies. The whole engine is public/verifier.js.

For AI agents — MCP server

The same engine ships as a local, keyless MCP server so an agent can verify its own citations before it emits them (hallucinated citations are caused by LLMs — catch them in the loop). All registry calls run on the user's machine; there is no hosted backend.

cd mcp && npm install
# then point Claude Desktop / Claude Code / Cursor at:  node /abs/path/mcp/server.js

Exposes a verify_citations tool. See mcp/README.md.

Test

node test/run-tests.mjs        # 25 live-API cases: pass, fail, honesty, adversarial
cd mcp && node test-client.mjs # end-to-end MCP protocol test

Deploy

npx wrangler deploy       # assets-only Cloudflare Worker; see RUNBOOK.md

Docs

• DECISIONS.md — why this validator, this architecture, this design
• NOTES.md — build log
• VERIFICATION.md — evidence the live deployment works
• RUNBOOK.md — operate, deploy, roll back, teardown

MIT licensed. Built on free public scholarly infrastructure; not affiliated with any registry.