---
title: "NLWeb, WebMCP and MCP: How Agents Talk to a Website"
url: https://hostmy.blog/nlweb-webmcp-mcp/
date: 2026-10-26
modified: 2026-10-26
lang: en
author: "Aditya Sharma"
description: "NLWeb, WebMCP and MCP are three different things. A plain comparison of what each is, where it runs, and which one a WordPress site should care about."
categories:
  - "RankReady"
image: https://hostmy.blog/wp-content/uploads/2026/10/hmb-card-2626-1024x538.jpg
word_count: 1351
---

# NLWeb, WebMCP and MCP: How Agents Talk to a Website

NLWeb, WebMCP and MCP are three names for three different things, and they are routinely mixed up. All three are about the same hope: that an AI agent could deal with a website through a clear interface instead of guessing from its pages. They differ in where they run, who they are for, and how far along they are.

This post compares them as protocols. It does not tell you how to set any of them up on WordPress. It is aimed at the site owner who keeps hearing the terms and wants to know which, if any, to care about. The two explainers it builds on are [WordPress MCP: what the Abilities API lets an AI do](/wordpress-mcp/) and [what WebMCP is, and whether a WordPress blog needs it](/what-is-webmcp/).

## The three in one table

| | MCP | WebMCP | NLWeb |
| --- | --- | ------ | ----- |
| What it is | An open-source standard for connecting AI applications to external systems | A proposed browser API for pages to offer tools to agents | An open-source project for a conversational interface to a site |
| Where it runs | On a server that an AI application connects to | In the visitor's browser, in the page's own script | On a server, as a service that answers questions about the site |
| Who it serves | AI applications and their users | An agent working in a browser tab | People and AI agents asking a site questions |
| Status | Supported across many AI clients and tools | A draft report from a W3C community group, not a standard | A public project whose README calls its samples proofs of concept |

The rows are the whole story. They solve related problems at different layers, and they are not rivals.

## Why the names get confused

The overlap in the names is the main source of the muddle. WebMCP borrows the idea of MCP and moves it into the browser, so it shares most of the name. NLWeb is a separate project, but every instance also acts as an MCP server, so it shows up in conversations about MCP as well. And a site's own read-only manifest, such as the one RankReady serves, borrows the same vocabulary without being any of the three.

The way to keep them apart is to ask where each one runs. If it runs on a server and an AI application connects to it, think MCP. If it runs in the visitor's browser, think WebMCP. If it is a service that answers questions about the site, think NLWeb.

## MCP: the plumbing

The Model Context Protocol describes itself as an open-source standard for connecting AI applications to external systems. Its documentation compares it to a USB-C port for AI applications: a standard way to plug an AI into data sources, tools and workflows. In its [introduction](https://modelcontextprotocol.io/introduction), it lists assistants such as Claude and ChatGPT, and tools such as Visual Studio Code and Cursor, among the clients that support it.

On a WordPress site, MCP is the connection between an AI application and the site's abilities, which is what the adapter provides. It is about an AI doing things for you, with your login. Because it can change a site, it needs care, and [WordPress MCP security](/wordpress-mcp-security/) covers the limits to set.

## WebMCP: tools inside the page

WebMCP carries a similar idea into the browser. A page registers tools, each with a name, a description and a schema, and an agent working in that tab calls them instead of clicking around. The [draft report](https://webmachinelearning.github.io/webmcp/) comes from the W3C Web Machine Learning Community Group, and it says plainly that it is not a W3C standard and is not on the standards track. It runs in the visitor's browser and works while the page is open.

That makes it a natural fit for pages where people do things, such as booking or buying, and a poor fit for a page that is only text. The detail is in the WebMCP explainer linked above.

## NLWeb: a question-answering endpoint for a site

NLWeb is the odd one out, because it is a project and not a protocol. Its [README](https://github.com/nlweb-ai/NLWeb) describes it as simplifying the building of conversational interfaces for websites. The core idea is an ask method, where someone, or an AI agent, poses a natural-language question to a site and gets an answer back in Schema.org form.

Three details from the README are worth knowing. The README lists supported vector stores and language models, so running NLWeb means choosing one of each. It draws on the structured markup a site already has, mainly Schema.org on lists of things such as products, recipes, attractions and reviews, and it names RSS as a related format. And every NLWeb instance also acts as an MCP server, so it plugs into the plumbing above. The README also says its samples are proofs of concept and that most production deployments will build their own interface, and it marks some deployment options as still to come.

## How they fit together

| Question | The answer is | Because |
| -------- | ------------- | ------- |
| How does an AI application connect to a system? | MCP | It is the standard for that connection |
| How does a page offer actions to an agent in the browser? | WebMCP | The tools live in the page |
| How can a site answer questions in natural language? | NLWeb | It is a service built for that, and it speaks MCP |
| What data does NLWeb read? | Schema.org markup and feeds | That is what it is built to use |

The last row is the one that matters most for a blog. Whichever of these wins, the structured data your pages already carry is what it reads. Good structured data helps under any of them, and [structured data in WordPress: Article, FAQPage, Speakable and Person](/structured-data-wordpress/) covers writing it.

## What a WordPress blog should do about it

For most blogs, nothing here is a prerequisite for anything. Running NLWeb means running a service and choosing a vector store and a language model, which is a project and not a setting. WebMCP is a draft. MCP is useful if you want an AI to help you operate the site, which is a different goal from being found.

The useful work is the older list: let crawlers in, make sure the text is in the HTML, and write sentences that can be quoted. [AEO tools: what to measure, fix and skip](/aeo-tools-wordpress/) puts it in order. If your site publishes recipes, products or reviews, the one thing worth doing now is making that structured data accurate and complete, because it is the common input.

## Be careful with the phrase agentic web

The phrase is attached to a lot of selling. The honest state of things is that agents do not arrive at sites on their own through these protocols. RankReady's own WebMCP page says that discovery is not automated the way search crawling is, and that anyone selling it as an inbound traffic channel is ahead of the evidence. The same caution applies to the other two. Each is a real piece of engineering, and none is a source of readers you can count on today.

## Questions to ask anyone selling one of them

`1. Which one do you mean: MCP, WebMCP or NLWeb?
2. Where does it run, and what does it need from my site?
3. What can it read, and what can it change?
4. What evidence is there that anything uses it?
5. What does it cost to run, and what happens if I remove it?`

A clear answer to the second and third questions is the sign of a serious offer. A vague answer to the fourth is the sign of a pitch.

Last checked in October 2026 against the MCP introduction, the WebMCP draft report and the NLWeb README. All three are moving, so check the current versions.

## Where RankReady fits

RankReady serves a read-only manifest at /.well-known/mcp.json, which is its own take on letting an agent read a site, and it is off until you switch it on. It does not implement NLWeb or the browser proposal. It runs alongside Yoast, Rank Math, AIOSEO and SEOPress, and its schema output is one reason a site is easier to read. The free [AI Readiness Checker](https://hostmy.blog/plugins/rankready/ai-readiness-checker/) shows what an assistant can read on your live site today.

If an agent asked your site one question tomorrow, would your structured data give it a correct answer?