117 AI Tool Home Pages Mention MCP. 8 Publish a File an AI Agent Can Find: A Census of the Agentic Web

The idea of a "second internet" — websites that AI agents operate directly instead of reading — has moved from talk to proposals. Google's WebMCP lets a page register tools an agent can call in the browser (navigator.modelContext). Anthropic's Model Context Protocol (MCP) connects agents to a company's services. Google's Agent2Agent (A2A) protocol lets agents discover each other through an agent card. Each of them works by putting a machine-readable description somewhere an agent can find it.

That makes adoption countable. If a company has built for agents, there is usually a file to show for it. So we went looking, across every AI tool company in our directory — the companies most likely to be early.

The census

On 29 October 2026 we probed the 613 AI tool company websites in our catalogue that answered for the files today's agent protocols use, and read each home page for mentions of them. A file counted only if it came back as valid JSON with the fields its format requires — a web page or an error at the same address did not.

MCP has no single agreed discovery address yet; the two we checked are the ones proposed so far, so a company can run an MCP server without publishing either. That caveat matters, and we come back to it below.

What we found

117 of the 613 home pages (19.1%) mention MCP or the Model Context Protocol — most often an "MCP server" or "connect to Claude" feature. Eight publish an MCP manifest at a well-known address, among them Notion, Frase, Taskade and Scite. Three publish an A2A agent card, eleven an OpenAPI description. One registers WebMCP tools in its page: Hostinger. In total, 25 of the 613 websites publish at least one agent-facing file we could find.

The leftovers are as telling as the newcomers: five websites still serve manifests for ChatGPT plugins, a format OpenAI retired in 2024. Agent files, like any other file, go stale when nobody owns them.

Reading the gap honestly

The distance between 117 and 8 is not proof that 109 companies are exaggerating. Many MCP servers are distributed as packages, listed in registries, or documented on a developer site rather than announced at the root of the marketing domain. What the gap does show is that an agent arriving at these websites, with no prior knowledge, would find almost nothing to tell it what the company's software can do.

That is the stage the agentic web is at: plenty of announcement, very little discovery. WebMCP needs developers to implement new browser APIs; MCP discovery is still being standardised; A2A cards are rare even among AI companies. Readiness for agents is real work, but it is work for the next few years.

What is measurable now

The practical point is simple: the agent protocols are early, but AI answers are not. ChatGPT, Gemini, Perplexity and Google AI Mode already answer buyer questions about these very products, and whether they mention and cite a brand is measurable today.

1. If you have an MCP server or API, publish its description where agents look — and keep it current. A stale manifest is worse than none.
2. Don't wait for agents to fix discovery. The agents that will matter first are the AI assistants your buyers already use.
3. Measure your presence in AI answers now, so you have a baseline before agents start acting on them.

Measure the AI answers buyers already read

Track your visibility across AI platforms. Visibility Overview, in the AI Visibility Toolkit, gives an AI Visibility score with mentions, citations and cited pages over time, broken down by ChatGPT, AI Overviews, AI Mode and Gemini.

See what AI answers associate you with. Brand Performance shows how often AI answers connect your brand and competitors with specific strengths — the things an agent choosing a tool on a user's behalf would weigh.

Make sure agents can get in. Site Audit in the SEO Toolkit lists the pages blocked from AI search bots such as ChatGPT-User and OAI-SearchBot. An agent acting for a user is turned away by the same robots rules and firewalls as a crawler.

Find the questions agents will be asked. The Keyword Magic Tool filters keywords by intent and by the SERP features they trigger, including AI Overviews. The commercial questions that already get an AI answer are the first ones an agent will be sent to handle.

Bring the data to your own agent. Semrush itself now offers its data inside AI assistants — its home page invites users to "Ask AI. Get Semrush data." — which is the agentic web working in the other direction: your agent pulling the data, rather than an agent visiting your page.

The AI Visibility Toolkit has no free trial of its own; the 7-day free trial is Semrush One's, which includes it alongside the SEO Toolkit. To set your own AI baseline before paying, try Semrush One free for 7 days: its page offers "Unlimited access to all Semrush One tools" and notes "Offer available to new users only", checked on 28 October 2026. Those are Semrush's published terms, not something we tested; none of the data above came from a Semrush account.

Announced, not yet discoverable

Among 613 AI tool companies, one in five talks about MCP on its home page and one in twenty-five publishes anything an agent could find on its own. The agentic web is being announced faster than it is being built. Until discovery catches up, the AI layer that decides whether buyers hear about you is the one already answering their questions — and that one you can measure today.

How the probe worked

• Websites (613): the AI tool company websites in the TheToolsVerse catalogue whose home pages answered on 28 October 2026, probed once on 29 October 2026.
• Files: one request to each address in the table above, following redirects. A file counts only if it is valid JSON (not HTML) with the fields named in the table.
• Mentions: visible home-page text matched for "MCP" or "Model Context Protocol", "A2A" / "Agent2Agent", and "API". WebMCP: a navigator.modelContext call anywhere in the home page HTML, including inline scripts.
• What a file check can't prove: agent support published elsewhere — on developer subdomains, in package registries or MCP directories — or whether any file works. A mention of MCP can describe a client feature as well as a server. Script files loaded from other addresses were not read, so WebMCP use may be under-counted.

Written by Sohail Akhtar. The saved responses and probe script can be re-run.