CiteStamp MCP Server (com.citestamp/citestamp)

This read-only Model Context Protocol (MCP) server provides citation-health lookups over CiteStamp’s CC0 citation graph. It checks references against public scholarly registries (Crossref, DataCite, OpenAlex) and flags references that do not resolve to a known work or that have been marked retracted. The hosted server is remote.

🛠️ Key Features

Read-only MCP access to a CC0 citation graph
Reference checks against Crossref, DataCite, and OpenAlex
Flags non-resolving references (“do not resolve to a known work”)
Flags publisher-marked retracted items
toolCount: 5

🚀 Use Cases

Verify citations before an agent emits them
Detect hallucinated-citation failures involving missing or retracted works

⚡ Developer Benefits

Ground citation outputs in known works from public registries
Support workflows that reduce hallucinated citations and retraction-related errors

⚠️ Limitations

No local build or run; the server is hosted remotely
Read-only operation (no data creation or updates)

CiteStamp MCP server

Ground citations before your agent emits them.

CiteStamp is a free citation-health tool for researchers: it checks references against the public scholarly registries (Crossref, DataCite, OpenAlex) and flags the ones that do not resolve to a known work or that a publisher has marked retracted, the hallucinated-citation failure mode LLM users keep hitting.

This repo is the install guide for the hosted MCP server over CiteStamp's CC0 citation graph. There is nothing to build or run locally; the server is remote.


Endpoint	`https://mcp.citestamp.com/mcp`
Transport	streamable-http
Protocol	`2025-06-18`
Auth	None. Anonymous at 10 requests/min per IP. An optional free key (email info@citestamp.com) raises that to 60/min; Pro to 600/min
Data	CC0 citation graph; human-signed and machine-inferred claims kept strictly apart

The tools (all read-only)

what_supports / what_refutes: typed edges bearing on a claim
what_cites / references_of: the citing and cited works around a reference
edges_about: edges on a topic
retraction_listing: whether the Retraction Watch database lists a DOI as retracted, from the dated export CiteStamp joins weekly. Not listed means not in that export; it does not check the publisher record

Every result carries its tier: accountable edges are signed by a human identity (ORCID); inferred edges are machine-derived. Your agent can ask for signed-only.

Add it to your client

No key is needed. Every snippet below works as written.

Gemini CLI (extension)

bash

gemini extensions install https://github.com/citestamp/mcp

Or add the server by hand in ~/.gemini/settings.json:

json

{ "mcpServers": { "citestamp": { "httpUrl": "https://mcp.citestamp.com/mcp" } } }

Gemini Spark (custom app)

In the Gemini web app: Settings & help → Connected Apps → under "Custom apps for Spark" click Add a custom app → enter https://mcp.citestamp.com/mcp → Next. Type @ in a prompt and pick CiteStamp to make Gemini use it. Custom apps need access to Gemini Spark on a personal Google Account with Keep Activity on.

Claude (claude.ai / Desktop)

Settings → Connectors → Add custom connector → URL https://mcp.citestamp.com/mcp. Leave authentication empty.

Claude Code

bash

claude mcp add --transport http citestamp https://mcp.citestamp.com/mcp

Cursor (`~/.cursor/mcp.json`)

json

{ "mcpServers": { "citestamp": { "url": "https://mcp.citestamp.com/mcp" } } }

Cline (MCP settings)

json

{ "mcpServers": { "citestamp": {
    "type": "streamableHttp",
    "url": "https://mcp.citestamp.com/mcp" } } }

VS Code (`.vscode/mcp.json`)

json

{ "servers": { "citestamp": {
    "type": "http",
    "url": "https://mcp.citestamp.com/mcp" } } }

If a snippet drifts from your client's current config schema, your client's MCP docs win; the endpoint is the stable part.

Optional: raise the rate limit

A key is never required. If you hold one, send it as a header on the same endpoint:

code

Authorization: Bearer <your key>

For Claude Code that is --header "Authorization: Bearer <your key>"; for the JSON configs, add "headers": { "Authorization": "Bearer <your key>" } to the server entry. A bad key is a loud 401, never a silent downgrade to anonymous, so add the header only once you have a key.

What this is not

CiteStamp reports what public registries say at a dated point in time. It does not judge the truth or quality of any work, and its only verdict vocabulary is "does not resolve to a known work" and "marked retracted by the publisher". Statements about works, never about people.

identifier	string	required	DOI, arXiv id, OpenAlex id, or internal claim id
asserted_only	boolean	optional	If true, restrict to human-signed (asserted) edges. If false (default), return both asserted and machine-inferred edges.
limit	integer	optional	Maximum edges to return per tier (default 100, max 1000). Well-cited works have thousands; the response reports has_more and inferred_total so you can tell a page from the whole answer.
offset	integer	optional	How many edges to skip, for paging. Advance by the LIMIT you asked for, not by the number of results you got back — duplicates are removed after the window is taken, so a page can return fewer rows than it consumed. Keep going while has_more is true.

identifier	string	required	DOI, arXiv id, OpenAlex id, or internal claim id
asserted_only	boolean	optional	If true, restrict to human-signed (asserted) edges. If false (default), return both asserted and machine-inferred edges.
limit	integer	optional	Maximum edges to return per tier (default 100, max 1000). Well-cited works have thousands; the response reports has_more and inferred_total so you can tell a page from the whole answer.
offset	integer	optional	How many edges to skip, for paging. Advance by the LIMIT you asked for, not by the number of results you got back — duplicates are removed after the window is taken, so a page can return fewer rows than it consumed. Keep going while has_more is true.

identifier	string	required	DOI, arXiv id, OpenAlex id, or internal claim id
asserted_only	boolean	optional	If true, restrict to human-signed (asserted) edges. If false (default), return both asserted and machine-inferred edges.
limit	integer	optional	Maximum edges to return per tier (default 100, max 1000). Well-cited works have thousands; the response reports has_more and inferred_total so you can tell a page from the whole answer.
offset	integer	optional	How many edges to skip, for paging. Advance by the LIMIT you asked for, not by the number of results you got back — duplicates are removed after the window is taken, so a page can return fewer rows than it consumed. Keep going while has_more is true.

identifier	string	required	DOI, arXiv id, OpenAlex id, or internal claim id
asserted_only	boolean	optional	If true, restrict to human-signed (asserted) edges. If false (default), return both asserted and machine-inferred edges.
limit	integer	optional	Maximum edges to return per tier (default 100, max 1000). Well-cited works have thousands; the response reports has_more and inferred_total so you can tell a page from the whole answer.
offset	integer	optional	How many edges to skip, for paging. Advance by the LIMIT you asked for, not by the number of results you got back — duplicates are removed after the window is taken, so a page can return fewer rows than it consumed. Keep going while has_more is true.

identifier	string	required	DOI, arXiv id, OpenAlex id, or internal claim id
asserted_only	boolean	optional	If true, restrict to human-signed (asserted) edges. If false (default), return both asserted and machine-inferred edges.
limit	integer	optional	Maximum edges to return per tier (default 100, max 1000). Well-cited works have thousands; the response reports has_more and inferred_total so you can tell a page from the whole answer.
offset	integer	optional	How many edges to skip, for paging. Advance by the LIMIT you asked for, not by the number of results you got back — duplicates are removed after the window is taken, so a page can return fewer rows than it consumed. Keep going while has_more is true.

CiteStamp