What this check tests
The audit looks for a live MCP server in at most five places, in this order: up to two links on your homepage that point to your own site and end in /mcp or /sse (or sit on a host starting mcp.), then <origin>/mcp, mcp.<domain>/mcp and mcp.<domain>/, using the scheme of the audited URL. The whole probe has a budget of the smaller of three request timeouts and 15 seconds. Redirects are followed only for 307 and 308, to the same site, up to three times.
At each address it sends, in turn:
- A POST of
server/discoverwithMCP-Protocol-Version: 2026-07-28. A JSON-RPC result that listssupportedVersions, or one of the modern error codes (-32020, -32021, -32022), confirms an endpoint. - Unless step 1 got a redirect other than 307 or 308, or a server error (5xx), a POST of a legacy
initializerequest (protocol 2025-11-25). A 2xx result with a namedserverInfoconfirms an endpoint. - If that was a 400, 404 or 405, a GET asking for
text/event-stream. A first SSE event namedendpointmarks the old HTTP+SSE transport.
A 401 at step 1, 2 or 3 only counts when its WWW-Authenticate: Bearer header carries a resource_metadata parameter. A bare 401 does not register an endpoint. A 401 at step 3 registers the endpoint as HTTP+SSE. The first address that answers wins.
| Result | Score |
|---|---|
| Streamable HTTP server found | Pass, 100 (the server name is shown when it reports one) |
| Endpoint answers 401 with that Bearer challenge (at step 1 or 2) | Pass, 75 |
| Only the legacy HTTP+SSE transport, including a 401 with that challenge at step 3 | Partial, 50 |
| Nothing found | Fail, 0 |
This check is never skipped, and it weighs 1.5 in Agent Protocols. It is also the signal the audit uses to decide whether the MCP authorization check applies.
Why it matters to agents
An MCP client that knows only your domain has to guess where a server lives. The MCP specification's transport page marks HTTP+SSE as deprecated and describes a single MCP endpoint for Streamable HTTP. The 2026-07-28 revision says servers must implement server/discover, and explains how a client tells these modern servers from older initialize-based ones; this check tries both, modern first. A 401 with a resource metadata pointer is how the authorization section tells a client where to start sign-in.
Be realistic about who needs this. An MCP server is a backend integration for sites that offer tools or data to agents, such as an API, a documentation search or a store. A content site without one fails the check, and that is a true reading rather than a defect. The probe also only tries the fixed paths above, so a server on another path is missed unless your homepage links to a /mcp or /sse address.
How to fix it
Only if you run, or plan to run, a remote MCP server:
- Serve the Streamable HTTP endpoint at
https://example.com/mcporhttps://mcp.example.com/mcp. Answerserver/discoverand, for older clients,initialize. - If you cannot use those paths, add a plain link to your endpoint on the homepage, such as
<a href="/api/v2/mcp">. The path must end in/mcpor/sse. - If the server needs sign-in, return
401withWWW-Authenticate: Bearer resource_metadata="https://example.com/.well-known/oauth-protected-resource", and publish that document. MCP authorization and OAuth discovery check the rest. - Move off HTTP+SSE if you can; it caps this check at 50.
- Check it from outside your network, since the audit sends requests from its own address.
Once found, MCP server card can describe the server before a client connects. Browser-side tools are a separate mechanism: see WebMCP declarative and WebMCP imperative. Background: agent servers.