How AI agents find your website’s MCP server
An MCP server only helps if an AI agent knows it’s there. There are half a dozen ways to announce yours, from well-known files to a tag in your page head, and they range from an accepted standard to a convention someone made up last year. Here’s which is which, what agents actually do today, and the cheap signals worth publishing anyway.
Updated
Why discovery is the hard part
An MCP server is a web address that AI assistants can use to ask you questions, send you leads and book meetings. Connecting to one is solved: the protocol is stable enough that ChatGPT, Claude, Perplexity, Gemini Enterprise and others all accept remote servers.
Finding one is not. When a buyer asks an assistant to “get a quote from a plumber in Denver”, the assistant needs a way to go from your domain, or your listing in search results, to the address of your MCP server, and to trust what it finds there. The MCP project has worked on that question since 2025, with several proposals along the way. Some were dropped, one was accepted, and plenty of blog posts still describe the dropped ones as current.
Every mechanism and its real status
As of November 2026:
| Mechanism | What it is | Status |
|---|---|---|
/.well-known/ai-catalog.json + Server Card |
A list of a domain’s AI services, pointing to a card describing each MCP server | SEP-2127, marked Final (Extensions Track). Card schema labelled experimental; AI Catalog format still being drafted. |
/.well-known/mcp/server-card.json |
A card at a fixed path on your domain | SEP-1649, an early draft. Closed in January 2026 and replaced by SEP-2127, which recommends against this path. |
/.well-known/mcp |
A manifest listing your MCP endpoints | SEP-1960, closed a day after it was opened in December 2025 as a duplicate of SEP-1649. Never part of the spec. |
<link rel="mcp"> and Link header |
A pointer to the server in your HTML or HTTP headers | Convention only. No SEP, and “mcp” isn’t a registered link relation. |
| llms.txt mention | A line in a Markdown file written for language models | llms.txt is a community proposal; mentioning MCP in it is plain text, not a format. |
| Official MCP Registry | A central index of public MCP servers | Run by the MCP project, in preview. Built mainly for marketplaces and aggregators. |
| WebMCP | Tools registered inside your web page for a browser agent | W3C Community Group draft; origin trials in Chrome and Edge. |
| DNS records | A TXT record such as _mcp.example.com |
An individual IETF Internet-Draft with no formal standing. SEP-2127 considered DNS and chose not to use it. |
| Schema.org | Structured data about the business and its actions | A real standard, but it has no MCP type. Helps agents understand you, not find your server. |
AI Catalog and Server Cards (SEP-2127)
SEP stands for Specification Enhancement Proposal, the MCP project’s process for changing the standard. SEP-2127 started life as SEP-1649 in October 2025 and moved to a pull request in January 2026 under a new process. The core maintainers voted to accept it in August 2026, and the MCP site now marks it Final on the Extensions Track. That makes Server Cards an official, optional extension: clients and servers can adopt them, and nothing breaks if they don’t.
It works in two steps:
-
The AI Catalog. A domain publishes
/.well-known/ai-catalog.jsonlisting its AI services. Each MCP entry has an identifier, the typeapplication/mcp-server-card+json, and either the card’s URL or the card itself. Clients doing domain-level discovery should fetch this file first.{ "specVersion": "1.0", "entries": [ { "identifier": "urn:air:example-plumbing.com:mcp:assistant", "type": "application/mcp-server-card+json", "url": "https://example-plumbing.com/mcp/server-card" } ] } -
The Server Card. A small JSON document describing one remote server: its name in reverse-DNS form, version, a description of up to 100 characters, and its endpoints in
remotes. The recommended place for it is your MCP server’s address plus/server-card, but it can live anywhere the catalog points, including another company’s domain.{ "$schema": "https://static.modelcontextprotocol.io/schemas/v1/server-card.schema.json", "name": "com.example-plumbing/assistant", "version": "1.0.0", "title": "Example Plumbing", "description": "Ask Example Plumbing about services and areas, request a quote or book a visit.", "websiteUrl": "https://example-plumbing.com/", "remotes": [ { "type": "streamable-http", "url": "https://example-plumbing.com/mcp" } ] }
That last point is good news for businesses using a hosted service. The SEP’s own example is a restaurant whose bookings run through a reservations platform: the restaurant lists the platform’s card in its own catalog, and the platform hosts the card.
Some deliberate limits worth knowing:
- No tool list. Cards leave out tools, resources and prompts. Servers change what they offer by user and over time, so clients are told to ask the live server instead.
- Advisory, not binding. Clients must not use a card for security decisions and should prefer what the live server reports, through the in-protocol
server/discoverrequest, where the two disagree. - Still settling. The SEP is Final, but the extension repository that holds the schema still calls itself experimental, says the
v1schema is pre-release, and gives the protocol revision as “draft”. The AI Catalog format lives in a temporary Linux Foundation working repository shared by MCP, A2A and others, and is still being shaped. Expect field changes.
Serve both files over HTTPS with open CORS and caching. The extension’s discovery document asks for these headers:
# On /.well-known/ai-catalog.json
Content-Type: application/ai-catalog+json
Access-Control-Allow-Origin: *
Cache-Control: public, max-age=3600
# On the Server Card
Content-Type: application/mcp-server-card+json
Access-Control-Allow-Origin: *
Access-Control-Allow-Methods: GET
Access-Control-Expose-Headers: ETag
Cache-Control: public, max-age=3600
It also says hosts should send an ETag and answer If-None-Match with 304 Not Modified.
The older well-known paths
You’ll still see /.well-known/mcp and /.well-known/mcp/server-card.json in tutorials, plugins and readiness checkers, so it’s worth being clear about where they came from.
/.well-known/mcp/server-card.jsonwas the path in SEP-1649’s first draft, with a card format that included tool lists and authentication details. The SEP moved to SEP-2127, and its discovery rules now say a card should not live under.well-known, because.well-knownis for site-wide metadata and the catalog already says where each card is./.well-known/mcpwas proposed in SEP-1960 as a manifest of a server’s endpoints, capabilities and sign-in methods. A maintainer closed it the next day as addressing the same problem as SEP-1649. It never had a spec, so files at this path follow whatever format each publisher picked.
If you already serve these files, there’s no harm in keeping them: some agents and tools look for them, and they cost nothing. Just don’t treat them as the standard, and add the AI Catalog alongside them.
Link tags and Link headers
A tag in your page head is the signal most likely to be seen by an agent that has already landed on your site, because it’s in the HTML the agent is reading anyway, no extra request needed.
<link rel="mcp" href="https://example-plumbing.com/mcp">
The same pointer can go in an HTTP response header, which agents that only fetch headers can see:
Link: <https://example-plumbing.com/mcp>; rel="mcp"
Be clear about what this is. There’s no SEP for it, and “mcp” isn’t in IANA’s registry of link relations. SEP-2127’s authors considered header-based discovery and set it aside, because the client has to fetch a page first. It’s a convention some sites and plugins use, and a reading agent can follow it, but no assistant documents looking for it.
llms.txt, schema.org and WebMCP
llms.txt is a proposal from Jeremy Howard, first published in September 2024: a Markdown file at /llms.txt with your name, a one-line summary and links to the pages that matter. It isn’t about MCP, but it’s plain text written for language models, so a sentence pointing to your MCP server is one more place a reading agent can find it.
# Example Plumbing
> Licensed plumbers serving Denver and Aurora. Emergency call-outs, water heaters, drain cleaning.
## For AI assistants
- [MCP server](https://example-plumbing.com/mcp): ask about services and areas, request a quote or book a visit. Streamable HTTP, no sign-in.
- [Services](https://example-plumbing.com/services): what we do and rough prices
Schema.org structured data describes your business, services and opening hours, and its potentialAction property can describe actions such as making a reservation. There’s no MCP type, so it won’t point anyone to your server. Keep it because it helps agents understand who you are and what you do.
WebMCP skips discovery altogether for browser agents: tools are registered on the page itself, so an agent working in the visitor’s browser sees them on whatever page it’s on. It’s in origin trials in Chrome and Edge. Our WebMCP guide covers it in detail.
The MCP Registry and DNS
The official MCP Registry is a central index of public MCP servers, backed by Anthropic, GitHub, Microsoft, PulseMCP and others. It accepts remote servers as long as they’re publicly reachable. You publish with the mcp-publisher command-line tool, and names are tied to something you can prove you own: io.github.yourname/… through GitHub, or com.yourdomain/… through a DNS or HTTP challenge.
Two caveats. It’s still in preview, with a warning that breaking changes or data resets may happen before general availability. And it’s meant to be read by marketplaces and aggregators, not by assistants directly; the docs say host apps should use downstream registries. It suits a SaaS company whose server many people will add by name. For a local business, it’s a low-cost listing with no proven pull yet.
DNS discovery comes up often. An individual IETF Internet-Draft proposes an mcp:// address scheme, a /.well-known/mcp-server manifest and a fallback TXT record like _mcp.example.com "v=mcp1; endpoint=…". Anyone can submit an Internet-Draft; this one has no formal standing, revision 03 expired in September 2026, and its paths differ from MCP’s own. SEP-2127 considered DNS TXT records and rejected them because they can’t describe servers that live at a path or port. Skip DNS for now.
What agents actually do today
The setup docs for ChatGPT, Claude, Perplexity and Gemini Enterprise all describe one flow: a person or admin pastes a server URL. None of them describes fetching a business’s AI Catalog, card or link tag while answering a question. Server Card support has started in code, with experimental implementations in the official Python and Go SDKs and a demo in the Goose client, but we found no major consumer assistant that documents automatic discovery.
That leaves two realistic paths today. A reading agent, one that fetches your pages to answer a question, may notice a link tag, an llms.txt line or a footer note and decide to use the server, if its product lets it. And a browser agent sees WebMCP tools directly. Discovery files are a bet that assistants will start checking domains for catalogs, which is exactly what SEP-2127 sets them up to do.
Checklist: publish the cheap signals
Each of these takes minutes, and none of them harms anything if no one reads it.
- Publish
/.well-known/ai-catalog.jsonwith one entry for your server, served as JSON withAccess-Control-Allow-Origin: *. - Publish a Server Card at your MCP address plus
/server-card, or wherever your provider hosts it, with a reverse-DNS name on a domain you own and a description under 100 characters. - Add
<link rel="mcp">to the head of every page, and the matchingLinkheader if your host makes that easy. - Mention the server in
/llms.txt, in one plain sentence saying what it does and that no sign-in is needed. - Keep or add the older paths (
/.well-known/mcpand/.well-known/mcp/server-card.json) if your tooling already produces them. Cheap, but not the standard. - Add WebMCP so browser agents see your tools on the page.
- Let AI agents in. Check that robots.txt, your firewall and any bot protection don’t block the user agents assistants fetch with, or none of the above gets read.
- Keep your schema.org data accurate: business type, address, hours, services.
- Consider the MCP Registry if you’re a SaaS company that people will add by name; verify your domain and publish under it.
- Check it from outside. Fetch each file with no cookies and confirm it returns 200, the right content type and the open CORS header. Our free AI-readiness scan checks the link tag, discovery files, llms.txt, structured data, WebMCP and crawler access in one go.
How Salesman Hal handles discovery
Once your site’s MCP server is on, Hal’s WordPress plugin and Cloudflare Worker publish /.well-known/mcp, /.well-known/mcp.json and /.well-known/mcp/server-card.json, all pointing at your hosted server, add <link rel="mcp"> to every page, and load Hal’s WebMCP script so browser agents see your tools. On WordPress hosts that answer /.well-known/ themselves, the plugin also writes the files to disk. For other sites, a pasted head snippet adds the tag and script, and you redirect the /.well-known/ paths to copies Hal hosts. Nothing is published until the server is on, so agents are never pointed at an address that doesn’t answer, and known AI agents fetching the files are logged as discovery visits in your dashboard.
FAQ
Is /.well-known/mcp a standard?
No. It was proposed in SEP-1960, which was closed the day after it was opened as a duplicate of the Server Card proposal. The accepted approach, SEP-2127, uses /.well-known/ai-catalog.json pointing to a Server Card.
Will publishing these files get me recommended by ChatGPT?
Not on their own. No major assistant documents reading discovery files to choose which business to recommend. They make it possible for an agent to find and use your server once it’s looking at your site, and they position you for assistants that start checking domains.
Which single signal matters most?
For agents that read your pages, the link tag, because it’s in HTML they already have. For the future, the AI Catalog and Server Card, because they’re the format MCP’s maintainers accepted. For browser agents, WebMCP. Publishing all three takes very little time.
My site builder can’t serve files under /.well-known/. What now?
Many can redirect those paths to files hosted elsewhere, and SEP-2127 lets a catalog point to a card on another domain. Whether every agent follows the redirect is untested, so add the link tag and an llms.txt mention too.
Should I list my server in the MCP Registry?
If you’re a software company whose server people will add by name, yes; it’s free once you verify your domain. For a local service business it’s optional: the registry is in preview and is read mainly by marketplaces, not by assistants answering buyers.
What about DNS records?
There’s an individual IETF draft for an _mcp TXT record, but it has no formal standing and MCP’s own discovery proposal decided against DNS. Skip it until something changes.
Sources
- SEP-2127: MCP Server Cards – HTTP Server Discovery, Model Context Protocol
- SEP-2127 pull request, modelcontextprotocol on GitHub
- MCP Server Cards (experimental extension): README, schema and discovery, modelcontextprotocol on GitHub
- SEP-1649: MCP Server Cards – HTTP Server Discovery via .well-known, modelcontextprotocol on GitHub
- SEP-1960: .well-known/mcp Discovery Endpoint for Server Metadata, modelcontextprotocol on GitHub
- AI Catalog, Agent-Card on GitHub (Linux Foundation working repo)
- Discovery: server/discover (draft), Model Context Protocol
- The MCP Registry, Model Context Protocol
- modelcontextprotocol/registry, GitHub
- The “mcp” URI Scheme and MCP Server Discovery Mechanism (draft-serra-mcp-discovery-uri-03), IETF Datatracker
- Link Relations registry, IANA
- The /llms.txt file, llmstxt.org
- potentialAction, Schema.org
- Getting started with custom connectors using remote MCP, Claude Help Center
- Add custom MCP server, OpenAI
Further reading
- MCP for business websites, Salesman Hal
- WebMCP: a practical guide for business websites, Salesman Hal
Make your MCP server easy to find.
Hal sets up the server, the discovery signals and WebMCP in about ten minutes. Free during early access.
Or run the free AI-readiness scan first.