Skip to main content

Grounded academic citation tools for AI agents

Search scholarly databases, format citations in 60+ styles, and verify bibliographies against academic records. CiteMe returns evidence and discrepancies — not unsupported verdicts.

Remote endpoint · Streamable HTTP

https://citeme.app/api/mcp

No API key, no account. Opening that URL in a browser returns 405 — the transport is POST-only by design, so point an MCP client at it rather than visiting it.

The three tools

check_references

Check references

The user already has citations and wants them verified.

Takes raw bibliography text. Returns per-reference match status, identifiers, and issue labels. This is the one to reach for after an assistant generated, extracted, or received a bibliography — including one it wrote itself.

search_citations

Search citations

The user needs to find a source.

Searches by topic, title, author, DOI, or ISBN and returns matching works already formatted in the requested style.

format_citation

Format citation

The user already has complete metadata and wants it styled.

Takes title, authors, year, and optional venue/DOI/pages fields and returns a reference plus in-text citation. No scholarly lookup; formatted results may be cached.

Reading a verification result

An agent that reports “no record found” as “fabricated” is wrong often enough to matter. Every result carries these labels, and the server ships the same guidance as MCP instructions so the calling model gets it too.

[OK]
A scholarly record was located and nothing was flagged on it.
[NOT FOUND]
No record was located in the databases below. This means the reference needs a human to look at it — it is not proof of fabrication. Grey literature, regional journals, and very recent work are routinely real and unfindable here.
[FABRICATED?]
A conservative hallucination risk signal. It is tuned to raise few false alarms rather than to catch everything, so it misses some fabrications and will occasionally flag a real source. Treat it as likely fabricated, never as proven.
[GREY LITERATURE]
Plausibly real, but the kind of work these indexes do not carry — theses, institutional reports, non-indexed regional journals.
[RETRACTED]
The record exists and was withdrawn. Surface it; do not treat it as a clean match.
[PARTIAL]
A record was located but the supplied fields did not fully line up with it.
[REVIEW]
A record was located and something about it still warrants a look — a diverging DOI, for instance. It is the catch-all for issues that none of the labels above describe.

One limit worth stating plainly: a match does not confirm the fields are correct. The checker reports a status, and field-level corruption frequently does not change it — a wrong year, journal, volume, or page range on a source that genuinely exists usually still comes back matched. Verify those by hand.

Quotas

The endpoint is anonymous, so the limits are per client and per IP rather than per account:

  • 20 tool calls per day per client — keyed by your MCP session id or app id when your client sends one. The transport is stateless and issues no session, so a client that sends neither is keyed by IP and this cap merges with the one below.
  • 50 tool calls per day per IP — a shared ceiling, counted across every client behind that address.
  • 120 requests per minute per IP for the transport itself, which covers handshake and tool listing.

Discovery is free: listing the tools never consumes the daily call budget. If you are behind a shared gateway or proxy pool, you may hit the per-IP ceiling before your own per-client cap — tell us and we will look at it.

What happens to your text

  • A bibliography you check is not published, not linked, and not attached to an account.
  • Results may be cached for up to six hours so an identical re-check is not re-billed against the databases.
  • A language model is used to split and parse what you paste — it never decides whether a source exists. That answer always comes from the indexes below.
  • Resolved works may be merged into CiteMe’s shared scholarly catalog.

Full terms in the privacy policy.

Databases queried

Verification is a lookup against these indexes, not a judgement produced by a model. A reference none of them carries cannot be confirmed here — it comes back as [NOT FOUND], or, when something about it stands out, as [GREY LITERATURE] or [FABRICATED?].

  • CrossRef
  • OpenAlex
  • PubMed
  • Europe PMC
  • Semantic Scholar
  • arXiv
  • Google Books

Coverage details and per-source behaviour live on the data sources page.

Also available

REST APIThe same searches and lookups over HTTP, with an OpenAPI spec.Reference checkerThe same verification, in the browser, for a whole bibliography.Benchmark corpus340 labeled references — correct, corrupted, and fabricated — with methodology.Apps and extensionsChrome, Firefox, Edge, Safari, and the Obsidian plugin.