MCP server contentgate

mcp-server-contentgate

MCP server for pre-deploy content checks. Catch the problem while the page is being written, not after it is live.

Documentation and background: adsandcode.com/en/open-source/contentgate

{
  "mcpServers": {
    "contentgate": { "command": "npx", "args": ["-y", "mcp-server-contentgate"] }
  }
}

Tools

Tool	What it answers
check_keywords	Do the terms this page is meant to rank for actually appear in it?
check_translation	Is this translation still structurally the same document as its source?
check_internal_links	Does this page link to anything on its own site that makes money, outside the boilerplate? Absolute links count only for hosts in site_hosts.
check_claims	Does this copy state outcomes as facts in a way regulated advertising rejects? A disclaimer counts only when its negation governs the claim itself.
sitemap_lastmod	Has this page genuinely changed, or is the build about to lie about it?
content_hash	Stable hash of a page with volatile blocks removed.
audit_page	Is this page readable without JavaScript, marked up with JSON-LD that parses, honest about its outbound links, and complete in its images and favicon?
compare_pages	How many sentences do these two pages share word for word?

Every tool is pure: content in, verdict out. No filesystem access, no network access, nothing written.

The guards are also importable as plain functions, and importing starts nothing: import { checkClaims, auditPage } from "mcp-server-contentgate".

Page audit

audit_page runs six checks on one rendered page and returns a verdict for each; pass checks to run fewer.

• visible_text: characters readable with JavaScript off, leaving out <template> and elements with the hidden attribute. Google renders JavaScript; many AI crawlers read only the HTML the server sent, so an article injected by a script is invisible to them. 1200 is a house default, not a search-engine number.
• structured_data: every JSON-LD block parses, the @type values you require are present (default Article, BreadcrumbList, Organization), and there is a dateModified. A required type is satisfied by its schema.org subtypes (BlogPosting and NewsArticle are Articles; Corporation, NewsMediaOrganization or a LocalBusiness is an Organization). A ProfilePage is not required to be an Article.
• outbound_links: external links without rel nofollow, sponsored or ugc. Google does not require any rel on ordinary editorial links (only sponsored on paid or affiliate links and ugc on user-generated ones), so do not add nofollow to a citation to make this pass: list the sites you cite in cite_ok. Hosts are resolved the way a browser resolves them, so backslashes and extra slashes do not disguise where a link goes; registries anyone can register under (co.uk, github.io, pages.dev) are refused in cite_ok.
• images: <img> without alt or width. alt="" is accepted.
• favicon: home page only; an icon <link> in the head, in a format Google Search lists (BMP, GIF, ICO, PNG, JPEG, PPM, TIFF), read from the URL path or, for a data: URI, its MIME type.
• stock_phrases: an editorial style list, not a detector of machine-written text. The default list is one editor's Traditional Chinese taste and is meant to be replaced: connectives such as 因此， and 然而， flag ordinary sentences. Pass your own, and always for any other language.

compare_pages splits both pages into sentences (on 。！？ and on . ! ? before whitespace), ignores sentences under ten characters, and measures the shared share in characters in both directions.

Sources for each rule are in the repository README.

The one design rule

A guard that has never been shown to fail is decoration.

The server runs its negative tests at startup and refuses to start if they fail. A checker that cannot prove it still catches a known-bad input should not be answering questions about your content.

The same test file is run by the Python implementation in the same repository, and both are required to produce identical verdicts on every case.

Three behaviours worth knowing before you call it

Keyword matching never normalises whitespace. web hosting and webhosting are different strings with different search volumes. If you pass one, the other does not satisfy it.

Pass exclude_classes. A call-to-action block that appears on every page will otherwise satisfy check_keywords and check_internal_links everywhere, and both tools become decoration.

Set forbid_source_script: false when the target language shares the source script, such as Simplified Chinese from a Traditional Chinese source. Otherwise every document in that locale is reported as untranslated.

Python CLI

The same checks run as a deploy gate, reading your site from disk, in the contentgate package in the same repository.

Licence

MIT.

Maintainer

Built and maintained by Ads & Code.