Stateless MCP

Stateless MCP is a design approach within MCP, the common standard for connecting AI to external tools and data, where the server does not keep track of the conversation state with a client and instead handles each request independently.

Think of a call center. In the old model, one agent had to remember your case, so every time you called back, you had to be routed to that exact same agent. If that agent was busy, no one else could take over, and the wait got longer. The stateless approach moves that memory out of the individual agent and into a record kept outside the server, so any agent who picks up the phone can handle the request right away. That means you can simply add more agents to handle more calls at once.

The same logic applies to MCP servers. If a server has to hold onto the state of every connection each time a client connects, a single server can only handle so many connections. Removing that state means any server can pick up any incoming request, so you can handle more requests just by adding more servers.

The article describes the removal of statefulness as the core change in this MCP roadmap update, allowing servers to scale horizontally without holding state. This comes from SEP-2575, which eliminates protocol-level sessions and the initialization handshake. It's a separate improvement from Sessionless MCP (SEP-2567), which removed sessions, so the two shouldn't be confused.