what is citeguard fr?

CiteGuard checks whether the sources cited in a piece of writing actually back up the claims attached to them. This matters because AI-drafted text often includes citations that link to real, working pages but do not actually say what the text claims they say. CiteGuard was built to catch that gap. You give it a document in markdown or plain text, or a list of specific claim and URL pairs, and for each one it fetches the linked source, follows redirects, and falls back to an archived copy on archive.org if the link is dead. It pulls out the readable article text from the page, ignoring ads and navigation clutter, and supports PDF sources as well. Then a language model reads only that fetched text, never relying on its own general knowledge, and decides whether the source supports the claim, partly supports it, contradicts it, does not mention it at all, is too unclear to judge, or simply could not be reached. Every verdict comes with a quoted passage from the source so a person can double check it quickly, plus a confidence score, and a full document scan returns an overall citation integrity score out of 100. CiteGuard can be used three ways: as an MCP server that plugs into Claude Code, Claude Desktop, or similar AI tools, as a command line tool installed through npm, or through a free hosted web API with a daily request limit. The judge model is not fixed to one provider, it works with any OpenAI-compatible chat endpoint, including presets for a few common providers. The tool is explicit about its limits. It does not detect whether text was written by AI, it does not declare something objectively true just because a source supports it, and it does not search for better sources on its own, only checks the ones it is given. It is written in JavaScript, installable through npm, and released under the MIT license.