Skip to content

Compare

Which agent standard does your site need?

These four don't really compete. They work at different layers, and several can run side by side.

Side-by-side comparison

Webserver For AI

What it is
A web server module that makes every page talkable
Agent can read pages
Yes: guide, text and exact views of every page
Agent can take actions
Yes: forms and links become actions, with confirmation for risky ones
Changes to your app
None: install the module, configure it
Running cost
Runs inside Apache, no AI model on your side
Works with Claude / ChatGPT
Yes, every site is an MCP server
Status
Released (v1.5.0), open source

llms.txt

What it is
A static Markdown reading list at /llms.txt
Agent can read pages
Only the pages you list
Agent can take actions
No
Changes to your app
Write one file
Running cost
None
Works with Claude / ChatGPT
Only if the assistant reads it
Status
Community proposal

NLWeb

What it is
A service that answers natural-language questions about your content
Agent can read pages
Through answers built from your structured data
Agent can take actions
Through MCP tools you build
Changes to your app
Structured data (Schema.org) and a running service
Running cost
Hosting plus AI model costs
Works with Claude / ChatGPT
Yes, via MCP
Status
Open source, adopted by several platforms

WebMCP

What it is
A browser API: a page registers tools an in-browser agent can call
Agent can read pages
Not its focus
Agent can take actions
Yes, the tools the page's JavaScript registers
Changes to your app
JavaScript in each page
Running cost
None on the server
Works with Claude / ChatGPT
Only agents running inside a supporting browser
Status
Experimental browser trial

When ai-chat/1 is the better choice

Existing server-rendered sites

PHP, WordPress, Rails, Java or static sites gain readable pages and working forms without a new API or a JavaScript rewrite.

People and agents together

The agent uses the visitor's own session, so someone can log in or pick a product by hand, then ask the agent to finish the job.

No AI bill on your side

Page guides and actions come straight from your site. The visitor's assistant does the thinking and pays for it.

Real transactions, safely

Forms become actions with clear fields and error messages, and payments or deletions always ask for confirmation first.

Plain HTTP

Every URL stays one address: HTML for people, a conversation for agents, picked by the normal Accept header.

When to use each

Use llms.txt

You run a content site and only want to tell AI tools what to read. Cheapest option, easy to undo. Webserver For AI already generates one if you don't have it.

Use NLWeb

You have rich structured data (products, recipes, articles) and want a conversational search box, and can pay for hosting and model use.

Use WebMCP

Your site is a JavaScript app and you want agents inside the visitor's browser to press buttons you define. Expect it to change while it is experimental.

Use Webserver For AI

You run Apache and want agents to browse, ask and act on your existing site, logins and forms included, without rewriting the app or running an AI model.

Combining them

The module serves /llms.txt and an MCP endpoint, so it already covers the reading-list layer and the assistant-connector layer. An NLWeb service or WebMCP tools can sit alongside it on the same site.

Sources: llmstxt.org, nlweb.ai, NLWeb vs llms.txt vs WebMCP (Nov 2026). Try ours in the live demo or read how it works.