Agentimus – AI SEO, llms.txt & MCP for AI Agents

Changelog

1.36.0

• New: the Dashboard says what your site runs. Four panels state every system’s standing in one look — the doors agents use, the signals you announce, what search shows, and what your writing holds — each led by one number, each ending in the link that opens the full screen. The numbers are the same ones the nav and the screens already show, read from the same payloads, so the card can never disagree with the report it points into; and a switched-off system is stated rather than hidden, because off is information.
• New: page-builder pages tell the truth now. On a page Elementor or Beaver Builder owns, the visible body never lived in the WordPress content — so the .md twin, the description fallback and the readability grade could all read a body no visitor sees, and an agent replacing the content was told “saved” while the page did not change, right up until the builder’s next save quietly erased the edit. Now every machine surface asks the builder for the real rendered page, and an agent’s body replacement on a builder-owned page is refused with the reason and the fields that do work there — on Divi and WPBakery, where the layout lives inside the content itself, that same refusal is what keeps an agent from destroying the design. Bricks, Oxygen and SiteOrigin pages get the same write protection, and one filter teaches Agentimus a builder it doesn’t know.
• New: pages splitting one search are now a finding. The engine can only send a search to one page at a time, so when several of your pages keep appearing for the same query they take turns — and every turn a weaker page takes is a click the strong one loses. Agentimus reads this from the engine’s own rows: Findings names the heaviest splits, and Visibility → Search lists each one with the page that earns the click stated outright, the weaker pages each carrying one decision — point it at the winner, or set it aside on the worklist’s own ledger. Thin data never accuses a page, and a query one page already owns is a result, not a problem. Connected agents read the same list from read-search-opportunities.
• New: two more questions a connected agent can ask. read-findings hands an agent the same ranked front door the Findings screen shows — everything open across the site, ordered by what it costs, bounded at twelve rows — and read-audience answers who reached the site: people and machines counted apart and never summed, with your site’s analytics folded in when Google Analytics is connected. The readiness tool also declares everything it actually returns now — the content worklist and the set-aside ledger had been riding along undeclared, invisible to any client that trusts the declared schema.
• New: Findings — one front door for everything open across your site. The score could read 99 and “Excellent” while five separate screens quietly held work you had never seen. Findings gathers all of it into a single ranked list — pages losing a click they already earn, and anything the setup checks caught — orders it by what each one actually costs you, and gives every row the button that lands on the fix. It sits in the bar with its own count, and that count only ever measures work you can do something about: things that are merely waiting on a later report get a quiet dot instead, because a number in a nav is a promise that doing the work makes it go down.
• New: light and dark, and your device decides. The toggle shows the mode you are switching to, and dark is a hand-tuned palette rather than an inverted light one — with its own dialect under each WordPress admin colour scheme, so it sits with the rest of your admin instead of fighting it. A choice you make by hand holds until your system next changes; after that your device’s own rhythm wins again, silently.
• Changed: the search worklist moved to where its numbers live. Search Opportunities — the list that turns Search Console figures into to-dos — now sits directly under the search report it reads from, on Visibility → Search, instead of at the bottom of Readiness. Readiness keeps a pointer to it. How search went, then what to do about it, in one view.
• New: Bing counted daily, and asked live. Bing’s human traffic now carries a day-by-day series with a week-on-week trend and the site’s true totals — its query endpoints report weekly buckets, so the daily numbers come from the traffic-stats endpoint instead, and every door drives its totals from that series. You can also ask Bing about a single page on the spot and get back what its crawler knows about it.
• New: Analytics, on the key you already connected. Connect Google Search Console and you can read visitor numbers from that same key — readers, visits and page views over the window the dashboard reports, including the people the search engines sent. Connecting it fetches the first numbers right away instead of tomorrow.
• Changed: every data surface now says whether it is counting people or machines. The AI traffic screen is Visitors, its cards say “people”, and a page view is no longer called a read. The request log is the machine side, Visitors is the human side, the dashboard states both — and they are never added together, because they are not the same thing.
• New: set aside is a ledger. Anything you park is now listed and counted rather than merely absent, and every parked page travels with it — so “not meant to be cited” stays a decision you can review, instead of a hole in a list.
• New: your content types, and your veto over them. The settings card lists the types a plugin switched on apart from your own, and you can turn one off even when the plugin that added it wants it on. The writing assistant follows the same rule: a page gets no invented sections and no tags, because a page is not an article.
• Improved: every data screen re-reads when you come back to it. Leaving a screen and returning used to show the numbers from whenever you first opened the plugin. They refresh on return now, so what is in front of you is what is true.
• Improved: 100 is earned, never rounded into. A score of 99.6 reads 99 — the top of the scale means every check actually passes.
• Improved: the nav carries true names at every width, cards end where they should, every fold in the plugin looks and behaves like one control — including the index groups, which had been showing two carets and no pointer — and a card’s notes run the full width of the card while keeping a comfortable reading measure. A page title that links out is now coloured like a link, instead of matching the prose around it until your cursor reached it.
• New: see which tools an agent actually gets. Discovery counted 42 tools and listed only the 7 published to anonymous agents — the other 35 existed nowhere you could look. Every provider group now opens to show the tools it contributes, by name, and the screen states three separate things in three panels instead of running them together in one column: what a signed-in agent can run, the addresses it connects to, and what an anonymous one is handed.
• Improved: on Discovery, the MCP and Abilities addresses now say which door each one is, the way the API list already named its provider — you no longer have to recognise a door by reading its URL. Every well-known document now carries the standard it implements, a quiet served tick, a plain-words line saying what the file is for, and a link out — in columns that hold their alignment even when one address runs long. A real file on your server still wears the one loud mark, because that file wins.
• Improved: the Discovery words explain themselves where they stand. Capabilities says outright that its rows are permissions, not tools — what an API allows and who declares it — and points at the signed-in list for the tools an agent can run; the providers lead says what a provider is; and each count wears the same caption on the dashboard and on Discovery, so one number is never described in two voices. When several providers are held back for the same reason, the reason is stated once and the later rows defer to it.
• Improved: every expand-or-collapse caret is the same solid triangle the plugin’s folds draw. A font glyph at any size read as decoration; drawn geometry is all ink.
• Improved: the dashboard’s “worth knowing” note is one legible card: a lit bulb, then the two halves as labelled columns of plain sentences — and the machine half states two more facts that were always true: its log counts live, so today is already in, and agents are counted by the name they declare.
• Improved: the words assume less. A sweep through every screen for the owner who is not a web person: each section teaches its own vocabulary where it stands instead of assuming it, every pointer at a setting names a path that actually exists, and the score card explains its own name. Machine reader, crawler, endpoint and their kin now arrive with a plain sentence attached the first time each screen uses them.
• Fixed: clearing recorded visibility data asks first. The Clear button on the AI Visibility screen erased its records with one click and no question; it now confirms like every other destructive button in the plugin, and states what is about to be erased before it is.
• Improved: dimmed text in the dark palettes is a step brighter. The faint tier was tuned for atmosphere and had drifted below readable at the small sizes it mostly sets.
• Fixed: on a phone, a long address no longer folds into a one-letter-per-line column — provider endpoints and the connection addresses wrap as whole lines. The capabilities list also keeps clear air between its rows and its scrollbar.
• Fixed: a document is not a tool. Four of the abilities Agentimus registers are MCP resources — llms.txt, llms-full.txt, discovery.json and your agent card, documents an assistant attaches rather than tools it runs — and every count on screen was adding them to the tools. Discovery reported 42 agent tools where the site has 38, and the Agentimus group read “25 tools” for 21 tools and 4 documents. They are counted apart now, and the documents live in their own section — each linked to the file it serves, with a note saying the files are public even though this door is signed-in. The count now matches the MCP server’s own tally: eighteen read tools plus the five write tools.
• Fixed: your site’s own content API could go missing from its discovery document. wp/v2 is WordPress’s shared namespace — every custom post type registered for the REST API lands a route inside it — and one plugin describing a single route there was enough for Agentimus to decide core was already accounted for. On those sites the posts, pages, media and category endpoints disappeared from discovery.json and the API list while still answering every request that reached them. Now only something describing the namespace itself stands core down.
• Fixed: internal link suggestions land in the right place in the post editor on WordPress 6.5 and later, where the editor’s rich-text internals had changed underneath them.
• Security: the anonymous MCP handshake stops handing out sign-in-only tool signatures. Asking the server for its tool list without credentials returned all sixteen read tools with their descriptions and input schemas — including the one that names which sensitive paths the exposed-files check probes for. Those signatures are deliberately withheld from the published documents, and a test asserts it there; nothing was watching the protocol, so the same list left by the other door. initialize and ping still answer, so a client can still tell a server is here and a scanner never reads “401” as “no MCP”; the tool list now returns the 401 and WWW-Authenticate that point a client at the connect flow instead.
• Security: the MCP token is confined to its own route, so it can no longer be presented anywhere else in the REST API.

1.35.0

• New: Ask AI in the post editor, at three scopes. Rewrite or extend the block you are in; select several blocks and change them all with one instruction; or ask about the whole post and get back a list of proposed edits — rewrite this paragraph, delete that one, add a section here — each with the reason it was proposed, to accept or reject one at a time. Blocks the plan does not name are never touched, and every change steps back through the editor’s own undo.
• Changed: the writing assistant creates, the editor revises. Editing an existing post inside the assistant’s drawer is gone: that drawer could only safely handle the blocks it had written itself, so any post with a real-world layout was declined rather than mangled. Create draft now opens the post in the editor, where every block type is at home and Ask AI does the revising. Nothing you have written is affected — this removes a screen, not a capability.
• New: a page is not an article, and the assistant now knows the difference. It writes posts, pages and your own content types, and a page gets no invented sections, no image slots, and no tags or categories — where only you know a fact (an address, an opening time, a policy) it leaves a marked placeholder instead of writing something plausible. Choose the type on the brief screen and filter the list by it. Agents creating content over MCP follow the same rules, so a page drafted from your assistant reads like one drafted here.
• New: ask Google about one page, right now. Every problem row, and the page lookup, gains a Re-check that inspects that single URL there and then — one of the day’s 2,000 inspections, stored like any other — so a page you just fixed is never judged by yesterday’s answer. Connected agents can ask the cheap way too: read-google-index now takes a URL and answers from the stored checks, spending nothing.
• New: your site’s own documents, offered to be read rather than run. llms.txt, llms-full.txt, discovery.json and your agent card are now listed as MCP resources, so a connected assistant can attach them the way it attaches a file, instead of being told to fetch a URL. A document you have switched off is never offered — an agent should not be handed a promise that 404s.
• New: an agent can finally name a picture you already have. A search-media tool finds images by title and by alt text — the place a photograph is actually described — so an assistant setting a featured image can choose one from your library instead of inventing one. Gated by the same permission as uploading.
• New: cancel a check you started. A whole-site sweep runs as a chain of short requests and there was no way out of one but closing the tab. Cancel stops it after the answer already in flight (that inspection is spent either way), and the rest of the queue waits for your next press or for the daily check — it is not thrown away.
• Improved: the page lookup understands what a person types. “privacy-policy” and “yoursite.com/terms” mean the same page as “/terms/”; only genuinely foreign addresses are refused. Answering your own URL with “that address is not on this site” was the tool being pedantic about its own input format.
• Improved: images are suggested to fit the post. A short draft no longer gets the same number of image slots as a long one, and revision in the editor never caps the images a post already has.
• Improved: the AI Readability panel stopped scolding. “Thin content” is now “Not enough substance yet”, “Hard to read” is “Low reading ease”, and “Nothing concrete to quote” is “Short on specifics” — the same measurements, named as a stage rather than a verdict.
• Improved: text buttons across the admin stop underlining themselves — a light pill on hover, room reserved so nothing shifts under the cursor, and a real focus ring for keyboards.

1.34.0

• New: Clear Cloudflare’s cache from Agentimus — publishing or editing a post now clears its stale copies from Cloudflare’s edge (the one cache no caching plugin purges for you), along with the front page and the machine files agents read, and the edge panel gains a Purge button for everything else. Needs one extra permission on your token — Zone → Cache Purge → Purge; without it everything stays read-only, and the panel says so in words.
• Improved: In Google’s Index now reports your site, not a sample of it. The headline numbers cover every page checked — on a 200-page site with 100 pages out of the index, the first number you read says so — and the summary accounts for every page, so the healthy majority is quiet rather than unmentioned. The standing list of watched pages is gone: those green rows never changed, and which pages get asked about first is a schedule, which the card now states in a sentence instead of a list. What you see is what needs a look, or news.
• New: a page that Google has left out is now re-checked every day until it gets in — up to 20 a day, the one waiting longest first, rather than waiting its turn in the whole-site rotation. And when one finally gets in, the card says so for about two days instead of the problem quietly disappearing, which is the moment you were waiting for. Every problem page keeps its own door into Search Console, however many there are.
• Improved: a network blip during the index check no longer ends the run — a slow or missing answer from Google rejoins the back of the line and the check carries on, reporting a failure only after three misses in a row. The run resumes where it stopped either way, and finishes on its own if you close the tab mid-check, so one slow answer never costs a whole run’s inspection budget or asks for another press of Check now. Google’s inspection also gets the time it actually takes to answer.
• Improved: the robots.txt check now opens the file as each engine last read it — a door to Google’s robots report and Bing’s robots tester, one per connected source, available any day on demand.
• Fixed: the XML-sitemap readiness check still named the pre-1.33 legacy address while robots.txt advertised the standard one — it now names, and opens, the sitemap actually served.
• Improved: the transient “robots.txt changed” note no longer renders as a readiness row — a notice among checks read as a failing check and changed the report’s count. The readiness count stays fixed now; the change itself still reaches you in the weekly email.
• Improved: the automatic Cloudflare purge is polite about everything — a switch on the Cloudflare card turns it off, a permission refusal stands it down silently until a reconnect or a successful manual purge re-arms it, and where the server supports it the save’s response is closed before the edge is called, so publishing never waits on Cloudflare.
• Fixed: the page lookup on In Google’s Index kept showing its previous answer after the box was cleared or edited — an answer now clears with the question that produced it.
• Fixed: the sitemap-health line sat below the page lookup, where it read as a fact about the page you had just looked up. It now sits with the whole-site section it describes.
• Improved: the lookup hint says what the answer is, not only what it isn’t — your own record of what Google said, rather than a new question to Google.
• Improved: the markdown edition’s “never store this” cache instruction now speaks nginx’s own dialect too, closing the one cache layer that ignores the standard headers.
• Improved: Search Opportunities cards for pages with no post behind them — the homepage on many sites, an archive — can now be set aside like every other card. They are held by address, listed in the same visible set-aside list, restorable in a click, and the card says why it offers no editor buttons.
• Improved: a search card’s readability flag no longer waits for the page to be recently edited — every card’s page is graded the moment the card renders, so “Also in Optimize” and its Check readability door appear the first time you see the card.
• Improved: the search worklist’s fine print now names its terms — which title it means (the SEO title in the “Search & AI” box, or the post title when that field is empty), that set-aside covers both engines, and that a fixed page’s card clears when later reports improve, not when you press Save.
• New: IndexNow — publishing, editing or removing a post can announce the changed address to search engines the moment it happens: one standard ping that Bing (the index ChatGPT search and Copilot read), Yandex and other participating engines all listen to. Off by default, one switch on the Bing card; only the changed addresses are sent, and a key file at /.txt proves the pings are yours. Works for every public post type — and content that lives outside posts (a plugin’s own catalog tables) can join with one line: do_action( ‘agentimus_announce_url’, $url ).
• New: the Found by AI Search card answers “does Bing know my sitemap?” — Bing’s own record of your registered sitemaps, when it last read them and how many URLs they carried, in Bing’s own dates. The IndexNow status gets its own honest line there too.

1.33.0

• New: Classic search, measured — Search Performance and Search Opportunities screens for Bing Webmaster Tools and Google Search Console: what people searched, how often you appeared, and which pages sit one improvement from page one. Every number is the engine’s own, and automated probe traffic is named, never blended in.
• New: In Google’s Index — a daily watchlist plus a whole-site rotation through Google’s URL Inspection: every verdict in Google’s own words, problems grouped with Search Console deep links, any page’s answer one lookup away — plus week-on-week trend, Google Discover, and the health of the sitemap you registered.
• New: Found by AI Search — how much of your site Bing’s index holds and how cleanly its crawler gets in; ChatGPT search and Copilot read this index today.
• New: Cloudflare edge panel (opt-in, your own token) — what the edge answered or blocked before your server ever saw the request, with a warning when the edge disagrees with your crawler policy.
• New: a setup wizard on first visit, and an agent-ready front door — SKILL.md, auth.md, and an MCP handshake that answers strangers politely.
• Fixed: the promoted XML sitemap now serves at WordPress’s standard /wp-sitemap.xml, and the old addresses redirect to it — a moved sitemap had been silently failing search-console registrations.
• Improved: every data screen names whose behavior it counts — people or machines — and every external data pull is chunked or budgeted, resumes after failure, and reports what happened in words.