> instead of the usual every-server-for-itself chaos. The messy part isnt the per-server config, its that write tools are often too coarse. Central policy and audit dont save you if one MCP action still means "do the dangerous thing" with a giant prompt sitting in front of it, because the failure mode is still one bad tool call. I’d rather see servers expose tiny, typed write ops with hard preconditions than a new control plane that just records who burned the house down.