The Model Context Protocol (MCP) is shifting to a stateless architecture, effectively transforming MCP servers into standard, horizontally scalable HTTP services. By moving away from the constraints of sticky sessions and shared session stores, the July 28, 2026, specification release candidate systematically dismantles the operational friction that previously hindered enterprise-grade deployment. This structural change allows the protocol to bridge the gap between LLMs and local data without the overhead of stateful connections.
Prior to this update, deploying MCP at scale required complex infrastructure. The protocol mandated an initialize/initialized handshake and maintained session state via the Mcp-Session-Id header. This forced architects to implement sticky routing or rely on shared session stores to ensure that subsequent requests from a client reached the same server instance. For teams managing high-traffic agentic workflows, this was a persistent operational bottleneck that complicated load balancing and increased the risk of connection-related failures.
The July 28 specification changes everything by removing the protocol-level session entirely. Under the new model, every MCP request is self-contained. Protocol versioning, client information, and capabilities are now transmitted within the metadata of every individual request. Furthermore, the introduction of a new server/discover method allows clients to fetch server capabilities on demand. Because the state is no longer tied to a specific connection, any MCP request can now land on any server instance behind a standard round-robin load balancer. To support this, the streamable HTTP transport now requires Mcp-Method and Mcp-Name headers, enabling gateways to route traffic efficiently without needing to inspect the request body.
Beyond the core architecture, the update introduces a robust Extensions framework. By utilizing reverse-DNS identifiers and an extensions map, developers can now ship new capabilities-such as MCP Apps for server-rendered HTML UIs or the now-standardized Tasks-without requiring changes to the core protocol. This modular approach allows for independent versioning, ensuring that the core specification remains lean while the ecosystem gains the flexibility to evolve. It is a clear signal that the protocol is maturing from a research-focused experiment into a stable foundation for production infrastructure.
Stability is further reinforced by a new, formal feature lifecycle policy. Any feature marked for deprecation is now subject to a mandatory 12-month window before it can be removed. This policy applies to current changes, such as the deprecation of Roots, Sampling, and Logging, which are being replaced by more direct integrations like tool parameters, LLM provider integrations, and OpenTelemetry. These methods will continue to function for at least a year, providing teams with a predictable migration path rather than a sudden breaking change.
For those currently running MCP servers, the operational implications are significant. The removal of the handshake and session management means that existing deployments will require a migration to align with the July 28 spec. This includes updating SDKs to support the new requirements, such as the tightened JSON Schema 2020-12 standards. While this requires an upfront engineering investment, the long-term benefit is a simplified, more resilient architecture that aligns with standard web infrastructure patterns.
The community has framed this update as the moment that “MCP killed the session.” As Akshat Paul noted in his analysis on X,
The biggest change to the Model Context Protocol is here-MCP is going stateless.
This shift is widely recognized by developers as the largest revision since the project’s launch and is currently viewed as a high-priority migration target for production servers.
This transition marks a pivotal moment for agent infrastructure. By aligning with the operational norms of the broader web-statelessness, standard load balancing, and a clear, versioned lifecycle-the Model Context Protocol is positioning itself as a durable layer for the next generation of agentic systems. For architects who have been waiting for the protocol to reach a level of maturity suitable for high-availability environments, the July 28 release candidate provides the necessary framework to move from prototype to production.