MCP vs WebMCP vs NLWeb vs llms.txt vs A2A: which one does your website need?

Published October 30, 2026 · Last reviewed October 30, 2026

MCP and WebMCP both give an AI agent named tools to call: MCP serves them from a server you host, and WebMCP registers them inside your web page for an agent working in the visitor’s browser tab. NLWeb answers plain-language questions about your content at an /ask endpoint, llms.txt is a reading list of your pages, and an A2A agent card describes an agent that other agents can call. Most websites need one or two of the five, not all of them. On a business site, MCP’s first user is usually the owner, managing the site from an AI assistant. WebMCP still runs in Chrome and Edge origin trials, though Shopify has already switched it on for its Liquid storefronts.¹ llms.txt is cheap but rarely read: 97% of valid files got no requests in Ahrefs’ May 2026 data.² And for most services sites, none of the five on its own answers a buyer’s follow-up question and books the call. Below, as of October 30, 2026: where each one runs, who reads it today, how mature it is, what it costs to ship, and a decision table for four kinds of site.

These names keep turning up in one sentence, as if they competed for the same slot; they do not. This piece does not re-explain the architecture behind them; our three-layer agent stack does that, and which files AI agents actually read goes file by file.

We sell into this question, so here is the stake up front. Tobira Site Agent is an AI agent for your website that answers and sells to people and their AI agents. It comes up again in the MCP section, as an owner channel and in a demo customer account, and after the decision table, at the row it fills.

Five names, four jobs

Strip away the acronyms and four jobs are left. Each name does one of them.

1. List what is on the site. llms.txt is a Markdown file that points a language model at your most useful pages. It changes nothing about how the site works; it is a reading list.³
2. Answer questions about the content. NLWeb, an open-source project started at Microsoft, puts an /ask endpoint on a site that takes a question in plain language and returns structured results built from the site’s own data. The same instance also speaks MCP.⁴
3. Hand an agent tools it can call. MCP does this between an AI application and a server, wherever that server lives. WebMCP does it inside a web page: the page registers tools through document.modelContext, and an agent working in the same browser tab calls them instead of clicking through the interface.⁵⁶
4. Describe an agent so other agents can find and call it. An A2A agent card is a JSON file, usually at /.well-known/agent-card.json, that says what an agent does, which skills it offers and how to reach it.⁷

Reading the list that way settles most of the confusion. “WebMCP vs MCP” is not really a contest: they share an idea (tools an agent can call) and differ in where the tool lives and whose login it uses. MCP connects an AI application to a server you run, which the agent reaches over the network with its own credentials, wherever the agent itself is running. WebMCP runs in a tab a person has open, inside their signed-in session, with the browser between the agent and the page. Chrome’s own guide, by André Cipriani Bandarra, Alexandra Klepper and Kasper Kulikowski, compares MCP to a call center that works from anywhere, and then: “WebMCP is an in-store expert. It’s available only while the user is on your page.”⁸ “NLWeb vs MCP” is closer to a layering than a choice, because an NLWeb instance can expose itself as an MCP server. And llms.txt competes with none of them. It sits one level lower, as a map rather than an interface.

One job none of the five covers well: a live party on your side that holds the conversation, qualifies the request and takes it to a booking or a payment. That is the gap between a site agents can read and one they can address (agent-readable vs agent-addressable), and the decision table comes back to it.

Where each one runs, who reads it, how mature it is

The table below is the whole comparison on one screen. “Who reads it today” means evidence we could point to on October 30, 2026, with its date, not what a vendor hopes will happen.

The table shows three things.

First, the readers column is uneven. MCP has the broadest set of clients, but almost all of them are AI applications acting for a user who connected the server on purpose; a stranger’s assistant does not find your MCP server on its own today. WebMCP has the most direct path to visiting agents, but only inside specific browsers and trials. llms.txt is the cheapest file on the list and has the weakest evidence of being read outside documentation sites.

Second, maturity and adoption point in different directions. MCP and A2A have the most formal specs, both now governed under the Agentic AI Foundation at the Linux Foundation.¹⁹ llms.txt has the lowest bar to ship, and NLWeb is cheap only where a platform offers it. WebMCP sits in between: a real API with a draft spec, in browsers that enable it only for sites registered in a trial, directly or through their platform.

Third, the counts come from narrow samples. APIs.io probes API providers, not the open web, and Ahrefs’ sample skews toward technical sites, so no single number describes the whole web. Cloudflare’s scan of the 200,000 most visited domains in April 2026 found MCP Server Cards and API Catalogs together on fewer than 15 sites.²⁰ Kin Lane of API Evangelist, who reads the APIs.io data, puts the discovery problem plainly: “There is no single place to look, and there will not be.”¹⁴

MCP: tools for agents, mostly on the owner’s side

MCP is the most widely supported name on the list and the most often misplaced. It is a protocol between an AI application (the client) and a server that exposes tools, resources and prompts. The 2026-07-28 specification made the core stateless: each request carries what the server needs, so any instance behind an ordinary load balancer can answer it. Standard request headers let a load balancer route each call without reading its body, and multi round-trip requests cover flows that need a follow-up question; our stateless MCP explainer has the detail.⁵ That makes MCP cheaper to run at scale. It does not make it a front door for strangers.

Here is why. An MCP server is useful to an agent only after someone connects it, which today means a person adding it to their AI application or a developer wiring it into an agent. Discovery is in progress. The Server Cards extension (SEP-2127) is marked Final and defines a card that tells a client where and how to connect (identity, endpoints, protocol versions); the tools themselves are still listed live after connecting. Domain-level discovery is left to AI Catalog, written jointly by MCP and A2A community members and still a draft.²¹²² Until a visiting assistant can find your server without help and decide to trust it, MCP on a business site mostly serves two groups.

• The owner and the team. Connect your CMS, CRM or site agent to the AI tools you already use and run them from there. WordPress exposes registered abilities through its MCP Adapter, private until a site opts each one in.²³ Our own product page, under “Manage from your AI”, lets the owner set up the agent and update prices, services and rules from Claude, ChatGPT, Codex or another MCP client, or through an API.²⁴ That is an owner channel, and we describe it as one.
• Known partners and power users. A customer who wants their assistant to check order status or book through your API can connect your MCP server on purpose. Our product page shows this too, in a demo account: a signed-in customer connects their own Claude or Codex to a business’s agent with a personal key.²⁴ Wix goes furthest here: every Wix site is its own MCP server at /_api/mcp, and people using AI tools that support MCP can ask about hours, make a booking or buy items.²⁵

Wix shows where MCP can reach visitors: a platform runs the server for millions of sites. For a single B2B site, a public MCP server is real work (authentication, rate limits, a tool surface that will not leak what it should not) for an audience that has to find you first. The specification makes authorization optional, so securing the server is your job; it also tells AI applications to keep a person in the loop who can deny a tool call.²⁶

Verdict. Ship MCP when you have a clear owner workflow or a partner who will connect, or when your platform gives it to you for free. Do not build one in the hope that ChatGPT will stumble onto it. If you sell software, an MCP server for your product (not your marketing site) may be the most valuable item in this whole article, but that is a product decision, not a website one.

WebMCP: tools inside the browser tab, still in trial

WebMCP takes the MCP idea of named, typed tools and moves it into the page. A site registers tools through document.modelContext (for example “search the catalog”, “check availability”, “submit this quote request”) and an agent working in the same tab calls them directly, instead of guessing which button to press.⁶ The person’s session, cookies and login stay where they already are, in the browser. That is the main difference from MCP: no separate server, no separate sign-in, and the human is present. It also shifts trust: OpenAI’s Site tools documentation treats the tool definitions and results a website provides as untrusted content.¹⁵

The status needs care, because it moves every few weeks and is often overstated.

• Spec. A Draft Community Group Report of the W3C Web Machine Learning Community Group, dated October 27, 2026 when we read it. Not a W3C Standard, not on the Standards Track, and not governed by the Agentic AI Foundation.⁶ A draft charter for a W3C Agentic Web Working Group names WebMCP as the starting input for its site-capabilities deliverable; nothing has been approved.²⁷
• Chrome. An origin trial from Chrome 149 through 156, with a desktop extension to Chrome 162 filed on September 28, 2026. Sites have to register for the trial, directly or through their platform; it is not on for every site.¹⁶
• Edge. Edge 153 (September 10, 2026) lists its own WebMCP origin trial.¹⁷
• ChatGPT. OpenAI documents “Site tools” as its implementation of the proposed standard, in the built-in browser of the ChatGPT desktop app, for tools registered on the top-level page, depending on model and rollout. Normal confirmation rules still apply to purchases.¹⁵
• Commerce platforms. Shopify says its WebMCP tools have been live on every Liquid storefront since August 5, 2026, with nothing to install, and BigCommerce said on October 21 that eligible Stencil storefronts have WebMCP on by default.¹⁸ Shopify said in August that agent support was limited to Chromium-based browsers through an origin trial.¹
• Other engines. Mozilla’s recorded position is neutral (August 5, 2026); WebKit’s is oppose (June 11, 2026).²⁸

So WebMCP has something the other four mostly lack: a named agent reader on the visitor’s side. Its reach is still narrow, because that reader exists only in specific browsers and trials, even where a platform has already done the site’s part.

Platforms are lowering the cost. In August 2026 Cloudflare previewed a dashboard switch that gives sites on its network a WebMCP interface from the edge (our write-up).²⁹ Stripe documents WebMCP tools for its Checkout and hosted invoice pages: it sets a consequentialHint on payment-submission tools and tells agent builders to get the person’s confirmation before calling them.³⁰ That last detail is worth copying whatever you build: mark the tools that move money or commit the customer, and make the agent ask (why that gate matters).

Verdict. If your site has a flow agents already try to drive through the interface (search, configure, quote, book), WebMCP is the most direct way to make that flow reliable for visiting agents, and the effort is mostly describing forms you already have. Join the trial if you want to learn early; do not plan revenue on it yet. If you run a Shopify or BigCommerce store, check what your platform already exposes before you build anything. Our WebMCP adoption piece has the earlier usage data, and it does not suggest heavy traffic so far. WebMCP is early, not failed: browser APIs often spend a long time in trials.

NLWeb and llms.txt: answering questions vs listing pages

NLWeb turns a site’s own data into something an agent can query. Microsoft introduced it in May 2025 as an open-source project: an NLWeb instance answers a natural-language question at /ask with structured results built from schema.org markup, feeds and similar data, and the same instance acts as an MCP server at /mcp.⁴ Launch partners included publishers such as O’Reilly Media and Hearst, plus Shopify, Eventbrite and Tripadvisor. The code now sits under the nlweb-ai GitHub organization, with its last commits in June 2026, and a W3C Community Group for NLWeb has existed since October 2025 without a published draft.¹³ Wix added a one-click opt-in in March 2026, starting with premium domains in English-language markets; its help pages say it draws on schema markup, product feeds and RSS and does not change how search engines crawl or rank the site.¹²

What nobody can tell you is how many sites run it, or which agents call it. presenc.ai’s review, with a status as of September 2026, found that no public count of deployments exists and that any figure in the thousands is an estimate.¹¹ Its advice to B2B brands with small sites: NLWeb can wait until agent demand shows up in their own logs. Catalog-heavy sites (retail, marketplaces, travel, recipes) are the likeliest to gain. We agree with that reading. If your content is a large structured catalog, NLWeb is a sensible experiment, cheapest on a platform that offers it. If your site is twenty pages about a consulting practice, an /ask endpoint over those pages adds little that a clear page and good markup do not already give.

llms.txt is lighter still: a Markdown file that lists the pages you want a model to read, with a sentence on each. Version 2 of the proposal (August 2026) lets the file sit at any path and cover the pages under it.³ The evidence on readers is now fairly consistent. Ahrefs looked at 137,210 domains: 28% published an llms.txt file, and 97% of the roughly 38,000 valid files got no requests at all in May 2026, from bots or humans, in a sample that skews toward technical sites.² Google Search Central says Google Search does not use llms.txt or similar AI text files, and that creating them “will neither harm nor help”.¹⁰ Where llms.txt is read, it is mostly by coding tools on documentation sites: Mintlify reports about 10.8 million fetches of /llms.txt across the docs it hosts in August 2026 (vendor data).⁹

Verdict. Publish llms.txt if your platform generates it or it takes an hour, especially for documentation. Do not expect it to change how assistants describe you; no major assistant has confirmed reading it at runtime. Treat NLWeb as an experiment for catalog sites. Neither one answers a buyer’s follow-up question about your terms, books a call or takes a deposit. Our llms.txt verdict has the longer version of the evidence.

A2A agent cards, and UCP or ACP if you sell from a catalog

A2A is a protocol for one agent to hand a task to another. Its agent card is the discovery piece: a JSON file that names the agent, lists its skills and says how to call it. The latest stable release is v1.0.1 (May 28, 2026), with a v1.1.0 release candidate tagged on October 29, and the project has lived in the Agentic AI Foundation since August 17, 2026; the path /.well-known/agent-card.json is a permanent IANA registration.⁷¹⁹

Two practical points for a website owner. A card describes an agent, so it is only worth publishing if there is an agent behind it that can take a task; a card that points at nothing is worse than no card. And the cards in the wild are often broken: API Evangelist’s September 2026 analysis of 427 cards found 60% conformant (up from 15% in July), 29% failing a hard structural check and 8% still at the old /.well-known/agent.json path.³¹ The well-known path also holds only one card per domain, which is why the v1.1 release candidate points hosts with several agents to AI Catalog.³² Our explainer on how A2A agent cards work covers the fields.

If you sell from a catalog, two more names belong in the table, and they are about checkout rather than discovery.

• UCP (Universal Commerce Protocol) covers checkout, cart, catalog and order capabilities. Google’s Universal Cart, which began rolling out in the US on October 15, 2026, checks out on UCP across retailers in Search and the Gemini app, with the merchant as seller of record.³³ For services, UCP is behind: its lodging booking spec is a draft, and appointments exist only as a proposal (issue #922, opened October 9, 2026).³⁴
• ACP (Agentic Commerce Protocol), from OpenAI and Stripe, continues and now supports product discovery in ChatGPT. OpenAI stepped back from Instant Checkout in March 2026 and lets merchants use their own checkout; its developer docs call checkout on the merchant’s own domain the recommended approach.³⁵

Verdict. Publish an A2A card when you run an agent that can actually take tasks from other agents, and validate it. If you sell SKUs, follow your commerce platform’s lead on UCP and ACP rather than building either yourself; commerce platforms and payment providers are doing that work. If you sell services, neither is ready for your booking flow yet (services need a different front door), which is why the buyer still ends up on your site. We cover that shift in discover in AI, buy on your site.

The decision table by site type

Here is the recommendation for four kinds of site, in the order we would ship. “Now” means worth doing this quarter; “Watch” means track it and revisit; “Skip” means not for this kind of site today.

Read down the B2B and services columns and a pattern shows. The standards that are ready (llms.txt, MCP) mostly serve readers other than a buyer’s assistant: coding tools and the owner. The ones aimed at visiting agents (WebMCP, A2A cards, NLWeb) are early or thinly adopted. And the most common thing a buyer’s agent needs from a services site, an exact answer to a specific question followed by a booked call or a quote, is covered by none of the five on its own.

That is the last row, and it is the addressable layer from our definition of an agentic website: a live party on your side that people and their AI agents can ask. It qualifies the request and takes it to the result you want. You can build it yourself on top of the formats above, or buy it. Site Agent is one option: its product page says it publishes files for AI (llms.txt, agent.json and AGENTS.md) and lets visiting AI agents ask their questions and take the next step, with leads reaching the owner as a short summary rather than a transcript.²⁴ In this table, that fills the last row, and its files for AI include an llms.txt. It does not replace a WebMCP flow in your own checkout or an MCP server for your product, and a site that runs Site Agent can still add those.

Two cautions on using the table. It reflects October 30, 2026, and three of its rows (WebMCP, NLWeb, A2A) could move within a quarter. And none of these formats, alone or together, gets a site recommended by an assistant. They let an agent read you, query you or act on your site once it is there.

What to remember

• Four jobs, five names: list pages (llms.txt), answer questions (NLWeb), give agents tools (MCP on a server, WebMCP in the page), describe an agent (A2A card).
• MCP serves the owner first, plus partners and customers who connect on purpose. Visiting assistants do not find MCP servers on their own yet.
• WebMCP has a documented reader in the visitor’s browser (ChatGPT’s Site tools), runs in Chrome and Edge only through origin trials, and is already on for Shopify Liquid and eligible BigCommerce stores. Early, not failed.
• llms.txt is cheap and read mostly by coding tools, NLWeb is an experiment for catalogs, and an A2A card needs an agent behind it.
• For B2B and services sites, the missing piece is a live endpoint that answers, qualifies and books.

When a buyer’s assistant reaches your site next month, which of these will it find, and who will answer its follow-up question?

FAQ

What is the difference between WebMCP and MCP?

Both give an AI agent named tools it can call. MCP connects an AI application to a server you host, which it reaches over the network with its own authorization. WebMCP registers tools inside a web page through document.modelContext, so an agent working in the same browser tab calls them within the person’s own signed-in session. WebMCP is a W3C Community Group draft in Chrome and Edge origin trials; MCP is a versioned specification under the Agentic AI Foundation.

Can WebMCP replace an MCP server?

No. WebMCP tools exist only while your page is open in a visitor’s browser, and they act inside that visitor’s session. Background jobs, server-to-server integrations and AI applications that work outside the browser still need MCP or an API. Chrome’s own guide calls the two partners rather than opponents, and many sites will run both.

Is NLWeb the same as MCP?

No. NLWeb is an open-source project, started at Microsoft, that answers natural-language questions about a site’s content at an /ask endpoint, using schema.org data and feeds. Each NLWeb instance also acts as an MCP server, so MCP clients can query it. MCP is the general protocol; NLWeb is one way to put a question-answering service on a site that speaks it.

Do I need llms.txt if I have WebMCP or NLWeb?

They do different jobs. llms.txt is a list of pages for a model to read; WebMCP and NLWeb are interfaces an agent can call. llms.txt is cheap to publish, but the evidence says few agents read it outside documentation sites: Ahrefs found 97% of valid files got no requests in May 2026, and Google Search does not use it.

Which of these does a B2B services website need first?

An llms.txt file if your platform generates it or it takes an hour, an MCP connection for the owner’s own tools if you want to manage the site from an AI assistant, and a live endpoint that answers buyer questions and books calls. WebMCP is worth watching, an A2A card is worth publishing only once an agent behind it can take tasks, and NLWeb adds little to a small services site. As of October 2026, all three are early or thinly adopted for this kind of site.

Will publishing these formats get my site recommended by ChatGPT?

No format on its own does that. They let an agent read your pages, query your content or act on your site once it is there. Whether an assistant mentions or recommends a business depends on its own systems, which none of these files control.

Sources