MCP goes stateless in July 28 release candidate

The most important protocol in agentic AI is about to get easier to run at scale. TechCrunch reported that the Model Context Protocol's forthcoming July 28 revision drops stateful request handling for a stateless design, a change that sounds arcane and matters enormously to anyone operating agent services.

No handshake, no sticky sessions

The core move: the revision removes the initialize handshake and the Mcp-Session-Id header, so any MCP request can land on any server instance behind a standard load balancer. Today that is not possible. Arcade founding engineer Nate Barbettini told TechCrunch that "every one of those machines has to know about a session ID that some other machine handed out," a "serious pain" that forces sticky routing and shared session stores. Stateless handling eliminates both, letting operators scale MCP servers horizontally the way they scale ordinary web services.

The practical payoff is in operations. Sticky routing means a load balancer has to pin each client to the one server that holds its session, which complicates autoscaling, failover, and rolling deploys. A shared session store adds a database in the hot path just to keep those machines in sync. Drop the session and both problems go away: add a server, remove a server, restart one mid-request, and nothing breaks. For anyone running MCP at scale, that is the difference between a bespoke stateful deployment and a boring, well-understood web tier that any operations team already knows how to run.

What else the release candidate changes

The stateless shift leads a dense changelog. Per the MCP release candidate, Tasks and MCP Apps become independently versioned extensions; six SEPs strengthen alignment with OAuth 2.0 and OpenID Connect; list and resource responses gain ttlMs and cacheScope caching metadata; and tool schemas move to full JSON Schema 2020-12. New Mcp-Method and Mcp-Name headers let gateways route requests without inspecting the body. The spec also opens deprecation windows, each running twelve months or longer, for Roots, Sampling, and Logging.

Standard, meet scale

MCP is already the de facto tool-calling standard for agents. The unglamorous work of making its core horizontally scalable lowers the cost of running agent services at production volume, which is where adoption has been bottlenecked. The final spec publishes July 28. This is the follow-up to our earlier coverage of the countdown; see our reporting on the spec nine days out and the stateless countdown for how the debate got here.

What this means for the agent economy: when the connective standard becomes cheaper to operate at scale, so does every agent service built on it. A stateless MCP core quietly removes one of the larger infrastructure taxes on running GaaS, and it does so without asking developers to abandon the protocol they already adopted.