CiteStamp

Tanpa kunci, tujuh alat hanya-baca: periksa apakah referensi mengarah ke karya yang benar-benar ada, dan lihat apa yang mendukung, membantah, atau mengutip sebuah makalah.

Server MCP Terhosting

npx add-mcp 'https://mcp.citestamp.com/mcp'

Terpasang ke Claude Code, Codex, Cursor, dan lainnya

Dokumentasi

Check before you cite

Language models invent citations. Everyone knows it, and the bibliographies keep shipping anyway. CiteStamp is the live, typed citation graph your agent queries before it emits a reference. Human-signed claims and machine-inferred coverage are kept strictly apart, so your agent always knows which kind of evidence it’s holding.

your agent CiteStamp the public graph what_supports, what_cites, … MCP or REST, key optional rate: 10 or 600 / min edges for this id signed: named, ORCID inferred: machine sweeps both tiers, each labelled claims, each with its tier + attribution, provenance, page has_more? raise offset signed and inferred are never blended your agent CiteStamp the graph a tool call key optional rate checked edges signed + inferred both tiers claims + tier + who, + page has_more? next page the two tiers never blend

The same tools answer over REST at api.citestamp.com/api/v1/ask.

Want to look at it yourself first? The same graph is a map you can walk in a browser, no key and no account.

What works today, and how keys work

This page describes only what’s live. The REST ask endpoints and the hosted MCP server both answer anonymously, with no key at all, at 10 requests a minute per IP, so your agent can ground a citation this minute. A key only raises the rate. A Pro token ($35/year, self-serve at /account) lifts every surface, including MCP, to 600 requests a minute, up to 20,000 a day. If you need more than that for a platform or an institution, write to info@citestamp.com and we’ll set it up by agreement.

Hosted MCP server

The graph speaks Model Context Protocol over Streamable HTTP. Point your agent at:

https://mcp.citestamp.com/mcp

No key needed to start. To raise your rate limit, send a key as a bearer token (Authorization: Bearer YOUR_KEY).

Seven tools. The five graph tools each take an identifier (a DOI in any common form), an optional asserted_only flag and an optional limit / offset window:

toolanswers
edges_aboutevery claim where the paper appears, either side
what_supportsclaims that something supports this paper
what_refutesclaims that something refutes this paper
what_citespapers whose reference lists include this one
references_ofwhat this paper itself cites
outside the graph
retraction_listingwhether the Retraction Watch database lists a DOI as retracted
check_referencewhether one reference resolves to a real work, before it’s cited

Every graph result carries a tier (asserted = human-signed, inferred = machine-derived), an attribution, and provenance. Set asserted_only=true when only name-on-the-record claims should count.

retraction_listing answers from the dated export CiteStamp joins weekly, with its credit line. Not listed means not in that export; it doesn’t check the publisher record.

check_reference takes a DOI, an arXiv id, a PubMed ID or the reference as written, and answers found, confirm (with the reason), not found in the named sources on a date, or couldn’t check. An agent has no browser, so this one runs on CiteStamp’s server. It reads Crossref, DataCite and doi.org (which answers for every DOI agency), plus PubMed when a reference gives a PubMed ID and no DOI. It stores nothing about the reference and never adds anything to the graph. A not-found isn’t proof that a work was never published. To check a whole reference list from a server, use the reference-check API, which needs its own key.

Answers are paged, and they say so

A well-cited paper has thousands of edges. Every response carries a page object (limit, offset, returned, has_more, and inferred_total where that number is exact), so a page is never mistaken for the whole answer. The default limit is 100 and the maximum is 1000; page by repeating the call with offset raised while has_more is true. Advance by the limit you asked for, not by the number of rows you got back: duplicate edges are removed after the window is taken, so a page can return fewer rows than it consumed. inferred_total is left out rather than estimated whenever there’s no exact count.

REST, if you prefer

curl -s https://api.citestamp.com/api/v1/ask/edges_about -H "Content-Type: application/json" -d JSON_BODY
where JSON_BODY is: {"id": "10.1007/s12225-019-9804-7"}

The rate ladder, in numbers: anonymous 10 requests a minute per IP; Pro token ($35/year) 600 a minute, up to 20,000 a day. Production traffic beyond that is priced case by case while the service is small: info@citestamp.com.

Why this beats string-matching a bibliography

  • Tier honesty. The API never blends human-signed claims with machine guesses, so your agent can quote “a researcher put their name on this” and mean it.
  • Typed edges. supports / refutes / extends carry direction and meaning a flat citation count can’t.
  • Coverage the Crossref-only tools miss. Four machine-inferred attributions, and every edge names its own. What each one is openalex-snapshot is bulk citation metadata projected from OpenAlex, which publishes it at source; it is most of what the API returns. openalex-extracted reference lists and zenodo-depositor-declared relations come mostly from CiteStamp’s own sweeps of SSRN and Zenodo metadata, so DataCite-registered work (datasets, preprints, software deposits) resolves here too. forrt-fred marks a small typed tier of supports and refutes edges: replication outcomes coded by the FORRT Replication Database, CC BY 4.0 per edge and credited on each row, not signed claims. Inferred edges are not independent of the metadata they were derived from, and a sparse answer means the sweeps have not reached a work yet rather than that nothing cites it.
  • No verdicts. CiteStamp records who said what. It doesn’t grade claims, so it can’t feed your agent someone else’s judgment dressed as data.
  • Stable, public, replayable. Signed claims live on an append-only public log; the whole record is CC0 and independently verifiable. The inferred graph CiteStamp builds is deposited too (see Bulk data below).

One caveat, because agents deserve honest data: the machine-inferred tier is millions of edges deep, while the human-signed tier is young and small. Today it’s entirely the founder signing claims about his own publications, disclosed as such on each claim page. The tier label on every result means your agent always knows which it’s holding.

Bulk data: download, don’t crawl

Building a dataset? Download these instead of walking the per-work pages:

  • The signed log: CC0, append-only, every line signed and chained to the one before it.
  • The inferred graph CiteStamp builds, deposited at Harvard Dataverse (CC0).
  • The openalex-snapshot citations on paper pages are projected from OpenAlex, which publishes them at source.
  • Questions about specific works: the MCP server and the REST ask API above (anonymous, rate-limited).

Service metadata

Machine-readable description at /.well-known/citestamp.json; a plain-text brief for LLM crawlers at /llms.txt and /llms-full.txt.