The latest Model Context Protocol specification removes protocol-level sessions from remote MCP servers, allowing consecutive requests to reach different instances behind an ordinary load balancer. The change eliminates the initialize handshake and Mcp-Session-Id header that previously encouraged sticky routing or a shared session store.
For cloud deployments, the simpler request-response model can reduce infrastructure that existed only to preserve MCP protocol state. AWS says it also makes serverless options such as Lambda a more natural fit. An optional server/discover operation remains available when a client needs to learn capabilities before invoking tools.
Stateless does not mean an application cannot remember anything. Product-specific conversation or workflow data can still live in a database or another application layer. The specification also replaces held-open streams used for server-initiated requests with a pattern where a server returns input_required and the client follows up in a later request. New headers support routing and throttling, while W3C Trace Context supports distributed tracing.
The trade-off is that stream resumability has been removed. Clients may need to retry interrupted operations, so tools that change data should be idempotent or carry a deduplication key. Existing clients also require a migration period. AWS recommends tracking protocol versions at the gateway and retaining old session infrastructure until legacy traffic has disappeared rather than switching every deployment at once.