The upcoming Model Context Protocol (MCP) release, slated for July 28, shifts the standard from a session‑based design to a fully stateless architecture, a move aimed at easing deployment on typical cloud platforms as AI pilots mature into production workloads.
Stateless design eliminates session bottlenecks
Earlier versions of MCP kept track of each client connection, requiring servers to maintain session state throughout an interaction. That model worked for developers running MCP locally, but it forced requests to return to the same machine, limiting scalability in multi‑server environments.
“The session‑based model made sense when MCP servers were local processes on a developer’s laptop. In production, it became an operational tax,” said Muskan Bandta, cloud associate at ZopDev. She added that infrastructure teams often heard that MCP “was not quite” able to scale like other cloud services. With the stateless redesign, the answer is now affirmative.
Under the new protocol, every request carries the context needed for any server to handle it independently. Applications that require continuity across calls must now store and retrieve that state themselves, rather than relying on the protocol to do so.
New features support modern cloud workflows
Beyond removing sessions, the release introduces a Multi Round‑Trip Requests (MRTR) mechanism. This allows a server to ask for additional input via a standard request‑response exchange before proceeding, eliminating the need for a persistent client‑server link.
Routable transport headers are also added, enabling API gateways to direct MCP traffic without deep packet inspection. According to Amit Jena, AI development manager at Kanerika, the headers reduce processing overhead, lower latency, and let enterprises apply existing routing, rate‑limiting, and security policies more efficiently.
The protocol now incorporates an updated authorization framework built around OAuth 2.1 and OpenID Connect, along with deterministic caching of tool and resource listings. The caching aims to improve large language model prompt‑cache hit rates, potentially cutting token expenses.
Related: Travel Tech I Refuse to Leave at Home
Legacy components such as Roots, Sampling, Logging, the older HTTP+SSE transport, and Dynamic Client Registration are slated for deprecation, though they will continue to work in this version and any other released over the next year.
Deprecating Sampling may have the most noticeable effect. “Sampling let MCP servers invoke the LLM through the client, which meant the server had a callback path into the model without owning that connection,” Jena explained. “Deprecating it means rebuilding that trust boundary.” The change shifts responsibility for calling model providers directly, altering network architecture, authentication models, and possibly billing flows.
Teams have a year to audit their reliance on Sampling. Jena warned that organizations that have not implemented their own sampling logic might be unaware of dependencies in third‑party MCP servers they consume.
While the code modifications are modest, identifying hidden session assumptions across gateways, deployment scripts, and monitoring dashboards can be time‑consuming.
The stateless approach could make AI workflows more portable and resilient across distributed environments. By exposing state management to developers, the protocol gives them finer control over how context is preserved and shared among tools, a shift that may encourage more innovative orchestration patterns.
Enterprises that built custom infrastructure around the previous session model will need to conduct thorough audits. The transition is unlikely to be seamless for those deeply entrenched in the older architecture, and the effort required to locate and refactor hidden dependencies could delay adoption.
