Could you clarify what are those "custom RPCs"? And the term "consumer of clustered caches" is not crystal clear either.

If you want to disable writes on certain nodes, you can use Security features.

If you don't want to run those commands on certain nodes, don't send them there. If you want a second line of defense, if the command is a VisitableCommand, its acceptVisitor() method calls visitor.visitUnknownCommand() (or maybe another one if you inherit from existing commands). You can insert a custom interceptor on top of the stack on nodes where you don't want to execute those commands and filter them there. If the command does not step through the interceptor stack, you will need to do the check in its perform method directly, and you'll probably need to initialize it with some components. Check ModuleCommandInitializer and its uses in the query module for examples.

Using shouldInvoke() wouldn't be too convenient, because InvocationContext does not provide you the address of current node.

Anyway, custom commands is a pretty advanced hacking into Infinispan and you could end up relying on non-public API or API that we don't keep too stable. Is there a reason for custom commands rather than using Distributed Executors? Could you share some details what you're doing?