What caught my attention is not the “stateless” label itself. It’s the quiet implication that a bunch of infrastructure people had been building around MCP sessions may have been solving a problem the protocol no longer wants to have. That feels healthy. Also slightly inconvenient, which is usually how real progress looks.
AWS is basically saying: if MCP no longer needs protocol-level sessions, then stop treating it like a chatty connection protocol with sticky routing and shared session stores bolted on. Route requests normally. Scale horizontally without pretending every server instance needs to remember the same conversation. Use Lambda if that fits. On paper, that’s a cleaner fit for the way most cloud systems already want to behave.
But I wouldn’t oversell the simplicity. The article hints at the real tradeoff, and that’s where the interesting work moves. If the protocol stops carrying session state, then application state, retries, tracing, idempotency, and all the annoying edge cases don’t disappear — they just get pushed into the layers around MCP. That’s fine, maybe better, but it’s not magic. The protocol got simpler; your system did not.
That’s also why the line “The protocol is stateless. Your application doesn’t have to be.” lands well. It’s a useful correction to the usual protocol-versus-product confusion. People tend to read “stateless” as “no state anywhere,” which is nonsense for real tools and agents. You still need continuity for multi-step flows; you just don’t need to weld it into the transport layer.
The part I’d watch closely is migration. AWS says older clients still need session infrastructure for a while, and that sounds right. If you have both stateless and sessionful MCP servers in the wild, the gateway becomes the awkward middleman. That’s where clean architecture meets ugly reality. I think that transition period matters more than the spec change itself, because specs are easy to celebrate and hard to deploy.
One small thing I appreciated: the article doesn’t pretend resumability and server-initiated behavior were free. Removing them forces client retries and makes idempotency more important. Good. That’s the kind of tradeoff that tends to get hand-waved away in AI-infrastructure marketing until someone loses a side effect.
Reference: Stateless MCP Removes Session Affinity Requirements for AWS Server Deployments
