What does a stateful MCP session enable and what is the key scaling tradeoff?

A **stateful MCP session** means the server remembers things between calls inside one connection. The client negotiates capabilities once during the handshake, gets a session id, and reuses it for every later request. That memory unlocks multi-step work. A database server can open a transaction in one call, run queries in the next few, then commit or roll back at the end. The transaction handle lives in the server's memory the whole time. A stateless model where every call starts fresh cannot do this safely, because each call would forget the open transaction. State also lets the server push to the client. It can send notifications when a resource changes, support subscriptions, and even ask the model to sample a completion through the client. The cost shows up in production. Because the state sits in one server instance's memory, follow-up requests must hit that same instance. That is called sticky routing, and it makes spreading load across many replicas harder. The server also has to clean up sessions when clients vanish, or memory leaks.

A stateful MCP session lets the server hold context across multiple calls in one connection. Capabilities are negotiated once at `initialize`, then a session id ties every later request to server-side memory. The headline benefit is multi-step operations. A database server can open a transaction, run several queries, then commit or roll back. The transaction handle persists in memory across calls. A stateless per-call model cannot guarantee those calls share one transaction. State also enables server-initiated traffic: change notifications, resource subscriptions, and sampling callbacks where the server asks the host to run a model completion. The production cost is horizontal scaling. Session state lives in one instance's memory, so follow-up requests must route back to that instance. This is sticky routing, also called session affinity. It blocks the stateless round-robin balancing that makes replicas trivially interchangeable. Teams must add affinity at the load balancer or externalize state to a shared store. Servers also need session cleanup, since clients that disconnect ungracefully would otherwise leak memory. Streamable HTTP softens this. It allows a stateless mode where each request is self-contained, so simple servers can scale flat and skip affinity entirely.