Guide
The files AI agents look for
Agents.json, MCP server cards and the rest. What each one is for, which of them anything actually reads today, and how much of it is worth your afternoon.
Why new files keep appearing
A reader wants a page. An agent wants an action: book the table, check the stock, start the return. A page can be read by a model and guessed at, but guessing is a poor way to spend somebody's money, so a second layer is being built where a site says what it can do rather than what it says.
That layer is young enough that the names are still moving. Below is what each of the main ones is for, and how much weight it currently carries.
agents.json
A description of your API written for software that has to work out which calls to make and in what order. An OpenAPI file lists the endpoints. This adds the sequence: to book, search first, then hold, then confirm, and here is what each step needs from the last.
It is a proposal rather than a standard. If you already publish an OpenAPI description, this is a small addition. If you do not, write that first, because every agent story starts there.
MCP server cards
The Model Context Protocol is the part of this with real adoption behind it. An MCP server exposes a set of tools and resources, and a model connects to it and calls them directly. No parsing, no guessing, no scraping your page for a price.
A server card describes what a server offers, so a client can decide whether it is the one it wants. For most websites this is not urgent. For anything with an account, a catalogue or a booking flow, it is the most direct route an agent will ever have to you.
Agent to agent
Protocols in this group handle agents talking to each other rather than to a model: how one introduces itself, says what it can do, and hands a task to another. It matters if you are building agents. It does not yet change anything about a website.
What to do this quarter
- Get the ordinary things right first. An agent reaches your site through the same crawlers everything else does, and a blocked crawler ends the story before any of this matters.
- Publish an OpenAPI description if you have an API and do not already publish one.
- Consider an MCP server if, and only if, an agent would need to do something on your system rather than read about it.
- Treat the file-based proposals as cheap bets. Add them if they take an hour. Do not reorganise a roadmap around them.
What we check, and what we do not
Our scan reads your site the way an engine does: crawler rules, structure, whether a model can describe your business from the page. It does not call your API or connect to a server, and it does not score you on files nothing has committed to reading. When a proposal here gains real adoption, it will be in the scan, and the methodology page will say when that changed.
Questions about agent protocols
What is agents.json?
A proposed file that describes your API in terms an agent can act on: which calls exist, what order they go in, and what each one needs. It sits alongside an OpenAPI description rather than replacing it.
What is an MCP server?
A small service that exposes tools and data to an AI model through the Model Context Protocol. Rather than a model reading your website, it calls your server and gets structured answers back.
Is agentic SEO a real thing?
The work is real, the name is doing a lot of lifting. It means making a site and its APIs usable by software that acts on a person’s behalf rather than by a person reading. The checks overlap heavily with ordinary AI readiness.
Do I need an MCP server for my website?
Only if an agent would need to do something rather than read something. A brochure site does not. A booking system, a stock lookup or an account tool is a different case, because an agent that can call it does not have to guess from a page.
Are these standards settled?
No. MCP has real adoption behind it. The file-based proposals are young, and some will not survive. Build on the parts that are already read by something, and treat the rest as cheap bets rather than requirements.
The ordinary checks come first.
Crawler access, structure and whether a model can describe your business. Free, in under a minute.