CiteProof MCP
MCP server with 40 tools that checks quotes, numbers, DOIs and PRISMA counts in AI-written literature reviews. No model inside.
- 110 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- The server reads the texts you pass and calls Crossref and doi.org for DOIs, computing the rest locally.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/citeproofDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The server reads the texts you pass and calls Crossref and doi.org for DOIs, computing the rest locally.
This is third-party code. Review the repository files before installing.
What it does
CiteProof adds evidence checks to a research agent. Language models invent quotes, figures and references, so the server compares each quote with the source text you supply, each number with its quote and looks up each DOI in Crossref and doi.org with a retraction check. There is no language model inside: searching, reading and judging stay with the agent, while the server stores decisions, checks evidence, computes numbers and refuses bad input. The server checks that PRISMA counts add up, computes Cohen's kappa for two reviewers, matches in-text citations to the reference list and checks overall risk-of-bias judgements against RoB 2, ROBINS-I and Newcastle-Ottawa rules. The repository also holds a separate web app with 24 agents, and the `citeproof-review` skill accompanies the server.
Who it is for. Researchers and students running systematic reviews with an agent who want to catch invented quotes and wrong numbers.
Good fit when
- The agent quoted a paper and you need to confirm the quote is in the source text.
- A data table holds sample sizes and confidence limits that must match the quote.
- A PRISMA flow diagram must agree with the recorded searches and screening.
- Two reviewers screened papers and you need an agreement coefficient.
- DOIs from the reference list need checking for existence and retraction.
Not a fit when
- You need the search and writing of the review itself; the server neither searches nor writes, the agent does.
- Work runs without access to Crossref and doi.org, so DOIs end up unverified.
- "Resolved" is read as "supported"; it means only a match with the reference list.
Example request
Check that every quote in the evidence table appears in its source text, verify the DOIs and show whether PRISMA counts add up.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
CITEPROOF_MCP_HOME- Folder for review data, ~/.citeproof-mcp by default.
Limitations
The server checks evidence but does not judge the scientific value of conclusions. An unreachable registry gives unverified, never not found. Resolved means a match with the list and not that the source supports the claim. The only network calls are DOI and retraction lookups to Crossref and doi.org. Review data goes to `CITEPROOF_MCP_HOME` or `~/.citeproof-mcp`. On first run the launcher package finds uv or installs it, which `CITEPROOF_NO_AUTO_INSTALL=1` turns off. In the Claude Science sandbox you may have to set `CITEPROOF_UV_HOME` inside the connector's workspace. It is tested on Claude Science, and other clients are still awaiting testing.
How to disable. Remove the connector in client settings or run `claude mcp remove citeproof`. Review data sits in `~/.citeproof-mcp` and is deleted by hand.
FAQ
How do I connect CiteProof to Claude Code?
Run `claude mcp add citeproof -- npx -y citeproof-mcp`. For other clients add the `npx` command with arguments `-y` and `citeproof-mcp` to MCP settings.
Do I need an API key?
No. The server has no language model inside, so no key is needed. The separate CiteProof Research web app does need a model.
How do I install it in Claude Desktop?
Download `citeproof.mcpb` from the latest release and open it under Extensions, or add the JSON with `npx -y citeproof-mcp` to `claude_desktop_config.json`.
What exactly does the server check?
Quotes against source text, numbers against quotes, DOIs through Crossref, PRISMA counts, reviewer agreement, citations against the reference list, risk-of-bias judgements and duplicates.
Where is review data stored?
In the folder from `CITEPROOF_MCP_HOME`, or in `~/.citeproof-mcp` when the variable is not set. Everything runs on your computer.
Discussion
No questions or reviews yet.
Related
MCP serversBibTeX verification against Crossref, OpenAlex, PubMed and arXiv: a CLI, Python API and MCP server to flag wrong references
Scholar Search MCP connects a discovered paper to its references and later citations when preparing a literature review.
Checks every reference of a paper against Crossref, dblp and arXiv before submission, with no language model in the verdicts.
A local MCP server for arXiv literature work: section-level paper reads, BibTeX from metadata, topic watches
A local researcher workspace for papers, evidence and drafts in one project, with AI proposals that people review and approve.
The MCP server of the seerai Zotero plugin gives outside agents 46 tools: library and scholarly search, reviews and a workspace.
README in short
The repository holds a 40-tool MCP server for evidence audit and a separate app for writing reviews with quality gates.