lurq-probe

What it requests

Per endpoint, per visit, at most these:

• One tools/list request. A stateless JSON-RPC call in the 2026-07-28 protocol revision. Servers on earlier revisions get the standard initialize handshake first, then the same tools/list, following nextCursor pages.
• OAuth discovery documents, only if the request was refused. The protected resource metadata (/.well-known/oauth-protected-resource, or the URL your WWW-Authenticate header names) and your authorization server’s metadata, in the order the MCP specification tells clients to try them.

What it never does

• It never calls a tool. No tools/call, no prompts, no resources read. It learns what your server offers, not what your server does.
• It never sends a credential. No API key, no token, no cookie. A server that requires one is recorded as requiring one.
• It never registers an OAuth client. Discovery reads your metadata documents; nothing is posted to a registration endpoint and no authorization flow is started.
• It never reaches a private network. Only public https addresses, checked again at connection time and on every redirect, so a listing cannot point it inside anyone’s network.

Which servers

On a schedule, only remote endpoints published in the official MCP Registry. An endpoint stops being probed once no registry entry lists it.

A developer can also paste a URL into lurq’s connect_check. That URL is probed once, on their request, and the result is not stored.

How often

• About once a day while your server answers, and every six hours for a few days after it changes.
• Less often when it does not answer: the gap doubles from six hours up to a week.
• A response that is not an MCP answer, including a rate-limit response, pushes the next visit out by days.
• At most two requests in flight to any one hostname at a time.

What lurq does with it

The result answers one question for developers: will this server work in the client they use, and what does connecting take? It records whether the endpoint answers, how clients sign in to it, deviations from the specification that strict clients refuse, and the tool names and schemas when they are readable without credentials. Accounts that run or approved a server are told when those change.

Stop the probe

Email contact@lurq.run with the hostname. Every endpoint on that host is taken off the schedule and not probed again. Removing the server from the registry has the same effect for that endpoint.

Something here does not match what you see in your logs? That is a bug on our side; the same address reaches us. For vulnerability reports, see Security.