Servers move, and nobody tells you
We have recorded 9,990 protocol era changes. Servers switch transports, gain auth, or quietly drop a capability — and the integration that depended on it breaks without a single error.
An MCP server is not a fixed thing. It gets redeployed, upgraded, put behind a login. We track which protocol era each server speaks on every round, and across the catalogue we have now seen 9,990 of those changes.
What a change looks like
Most transitions are small in appearance and large in effect. A server that answered your calls yesterday starts returning an authentication challenge today. One that spoke the older transport adds the newer one — or drops it. Nothing announces any of this.
- legacy to auth-gated: your calls now need credentials you do not have.
- dual to legacy: the modern endpoint you integrated against is gone.
- anything to unknown: the server is answering, but not in a shape we recognise.
Why your monitoring misses it
A server that starts demanding authentication is still up. It answers fast, it returns a valid HTTP response, and every uptime check you have goes on saying green. The failure only appears at the point where your agent tries to use it.
What helps
Watch the shape of the answer, not just its presence. If you depend on a server, record which era it spoke when you integrated, and alert when that changes — that single signal catches most silent breakages.
See recent era changes across the catalogue