Agentimus MCP Server

Checks ONE post/page’s readability for AI: grades how easily an AI can read, section and cite it — word count, an opening summary, concrete figures or cited sources, heading structure, quotable passage length, link density, image alt text, and freshness. Returns a pass/warn/fail row per check plus a tally. Use it when asked to check a page’s readability, or to tell an author exactly what to improve on a specific page.

Given an IP address, resolves who it really belongs to via forward-confirmed reverse DNS: the PTR hostname, the owning network/organisation, whether it maps to a known AI engine, and a verdict (0 = no engine match, 1 = forward-confirmed engine, 2 = forged engine hostname — an impersonator). Use it to answer "is this crawler that claims to be GPTBot actually OpenAI?".

Returns the plain-Markdown version of a post — the ".md twin" Agentimus serves to AI clients that prefer clean text over rendered HTML. Read-only preview of a single post; the site-wide index lives at /index.md.

Returns the JSON-LD @graph Agentimus would emit — for a specific post (pass post_id) or the site-wide identity graph (omit post_id / pass 0). Includes both the structured object and a pretty-printed JSON string. Read-only preview: it reflects the current draft, and changes nothing.

Returns real visits this site received FROM AI assistants (ChatGPT, Perplexity, Gemini, etc.) over a day range: totals, the per-assistant leaderboard, the landing pages they sent readers to, a daily series, and a diagnostic of unrecognised referrers. Optionally filter to one assistant and/or a landing-path prefix.

Returns the latest citation-check run (the Visibility screen's Citations tab): whether AI assistants mention and cite each tracked product, the overall visibility score and citation rate, per-product share-of-voice against competitors, and the trend across recent runs. Empty (hasData=false) until a run has completed. Read-only; it does not start a run (running one spends the site’s AI credits).

Returns what Cloudflare saw from AI crawlers BEFORE this server did, when the owner has connected a Cloudflare zone: per-crawler totals (requests at the edge, served from cache, reached the server, blocked at the edge, bytes out), per-company rollups, and any conflicts between the edge's observed behaviour and the site's declared AI policy — e.g. the edge blocking a crawler the policy welcomes. The request log only sees what reached the server; this is the missing before-the-server half. Returns connected=false when no zone is connected.

Returns whether Google's index holds this site's pages, when the owner has connected Google Search Console. Google's index is what AI Overviews, AI Mode and Gemini grounding read — the Google counterpart of "Bing's index is what ChatGPT search reads". Three tiers share Google's 2,000-inspections/day budget (there is no bulk index report): a WATCHLIST (homepage, busiest pages, newest posts — every answer in rows) checked daily; PROMOTED PROBLEMS — pages any check found unhealthy join the daily check (stalest first, capped at watched.promotedDaily) until they heal; and a WHOLE-SITE ROTATION walking every published URL in daily slices — healthy pages become the site counts, every problem (watched or not) appears in site.problems, and a page that just healed announces in site.healed for about two days before going quiet. On sites small enough the rotation covers everything every day; site.cycleDays states the honest cadence. site.problems ships a bounded, every-bucket share of the problem rows; to walk ALL of one state's pages, pass problemsState (a stateKey) and page with problemsPage — site.problems then holds exactly that page of 50, in a stable order, while every count stays complete (pages = ceil(site.problemStates[state] / 50)). Presence only, no traffic (read-search-performance has the traffic). Returns connected=false with empty rows when no key is connected.

Returns the site’s AI-visibility health: the blended 0–100 AEO/GEO score with its band (Findable, Readable, Trusted, Optimized, Cited pillars), the impact-ranked "do this next" action plan, AND the full list of readiness checks with their pass/warn/fail status and the fix for each. Use this to answer "how ready is my site to be found and cited by AI, and what should I fix first?".

Returns individual requests AI crawlers and agents made to this site’s discovery endpoints (llms.txt, the .md twins, JSON-LD, sitemaps), newest first and cursor-paginated. Each row has the endpoint, the detected agent, its user-agent, the owning network, and a verdict (0=unchecked, 1=verified real engine, 2=spoofed impersonator). Filter by agent, endpoint, network, verdict, or a user-agent prefix.

Returns the pages that already rank in classic search but under-earn — the worklist behind the Search Opportunities card (Visibility → Search). Two groups: "almostThere" (ranking 8–20, one improvement from page one) and "seenNotClicked" (already on page one, but a click rate well under THIS site's own page-one median — never an industry benchmark). Each page carries its searches, its totals, and whether it qualified on a single search or on the page's combined demand. Pair it with write-description / update-content to act on what it finds. Returns state "not_connected", "collecting" or "too_thin" when no honest verdict is possible.

Returns what classic search actually sent this site over the reported window: total times shown, visits, click rate and impression-weighted average rank, plus the top searches people used and the top pages they landed on. Numbers are the engine's own — Google Search Console and/or Bing Webmaster Tools, whichever the owner connected — never estimated, and never blended (the two count different searchers). Use it to answer "how is this site doing in search?". Returns source="" when no engine has reported yet. When source="bing", totals come from Bing's site-wide daily series (sample sums until that series has reported) while topQueries and topPages are its top-N sample counted separately — the lists sum to less than the totals, and one busy page can even out-count a tile. Neither is an error: state the split rather than reconciling the numbers.

Returns how much of this site sits in Bing's index and how cleanly Bing's crawler gets in, when the owner has connected Bing Webmaster Tools. Bing's index is what ChatGPT search reads today (Microsoft Copilot too), so this is the closest measurable answer to "can AI search find this site": pages in the index (daily trend), pages crawled, crawl errors, robots.txt blocks as Bing sees them, plus conflicts between Bing's view and the site's declared policy. Returns connected=false when no key is connected.

Returns the list of sensitive paths the exposed-files self-check probes for (config backups, .env, VCS metadata, DB dumps, keys) plus the site’s WordPress debug posture and detected environment. NOTE: this SUPPLIES what to check and flags the debug config — it does not fetch the URLs. The live probe must run same-origin from the admin browser so a server loopback cannot mask a leak the real public URL would reveal.

Finds images and other attachments ALREADY in this site's media library, so an agent can name one by id — most often as the featured_image on create-content or update-content, which accepts an attachment id but had no way to discover one. Searches titles, captions and descriptions AND alt text (alt is where photographs are usually described — a file called IMG_4831.jpg may carry the alt "sunrise over the river", and a title-only search would miss it). An empty query returns the most recent uploads, which is the right way to ask "what is in here?". Returns each attachment's id, title, alt, url, mime type, pixel dimensions and upload date — the url so a client that can see images may look before it chooses, the dimensions because a 300px logo is not a featured image. READ-ONLY: it uploads nothing and changes nothing. To bring in a picture the library does not have, pass an http(s) image URL as featured_image instead and the write tools import it. There is deliberately no way to upload bytes through this server.

Suggests which of the site’s OWN posts this post should link to, from local signals only (shared topics, categories/tags, and the candidate’s subject appearing in the text) — no AI call is spent. Each suggestion carries the target post, the exact phrase in this post’s text to link (empty when none exists — append a "See also" line instead), and a one-line reason. READ-ONLY: it suggests; to actually insert a link, edit the post through the governed update tool like any other content change.