A kill switch for MCP access

A stateless token is a good design right up to the moment you need it to stop. It validates on its own, which is what makes the gateway fast and what makes it survive a backend outage, and it also means that revoking the grant behind it changes nothing until the token runs out on its own clock. For an agent acting on your SaaS, that gap is the whole problem.
A revocation closes it. It refuses every token it covers and refuses to mint new ones, and it takes effect on the next call, within seconds. It is not limited to the panic case either: a block can be scoped to everyone in the tenant, or to one user, one agent, one server, or one client, so cutting off a single misbehaving agent does not mean stopping everybody's work.
It is meant to be lifted. A revocation is a state you can see and reverse once the incident is understood, not a configuration change somebody has to remember to undo.