One of the biggest practical limitations of today’s Model Context Protocol deployments has been state management. Every client session required initialization, servers had to keep track of session state, and production deployments often relied on sticky sessions or shared storage. The newly released Model Context Protocol 2026-07-28 specification changes that by making the protocol itself stateless, significantly simplifying infrastructure while expanding what developers can build.
Anthropic announced that Claude now supports the new specification, making it one of the first major AI platforms to adopt the updated protocol. The announcement also signals that the protocol has moved beyond experimentation and is becoming the foundation for production-grade AI tooling. Read Anthropic’s announcement.
The headline change is the removal of protocol-level sessions. Instead of negotiating a session during an initialization handshake, every request now carries the information needed for processing. That may sound like an implementation detail, but it has major operational consequences. Stateless servers can sit behind ordinary round-robin load balancers, requests become easier to cache and trace, and horizontal scaling no longer depends on session affinity. For organizations deploying fleets of MCP servers, that translates into a much simpler architecture.
The update also formalizes extensions as first-class parts of the protocol. Rather than treating advanced capabilities as loosely defined additions, features such as MCP Apps and the Tasks extension are now integrated into the specification. MCP Apps enable servers to expose server-rendered user interfaces, while Tasks standardizes support for long-running operations that extend beyond a single request-response cycle.
Security received considerable attention as well. The new specification strengthens authorization by aligning more closely with OAuth and OpenID Connect deployments. This is particularly important as MCP is increasingly used to connect AI systems with enterprise applications, where authentication and delegated access must fit into existing identity infrastructure instead of introducing custom mechanisms.
Another notable addition is a formal deprecation policy. Until now, developers had relatively little guidance on how protocol changes would evolve over time. The new policy introduces predictable migration windows, allowing SDK maintainers, client developers, and server operators to upgrade without unexpected compatibility breaks. Some older features, including Roots, Sampling, Logging, and the HTTP+SSE transport, are now deprecated with defined transition periods rather than disappearing immediately.
The protocol update also improves tooling around developer experience. The specification adopts full JSON Schema 2020-12 support for tool definitions, enabling richer validation and more expressive schemas. At the same time, beta SDKs for the major supported languages are already available, giving developers a migration path before the specification becomes the default across the ecosystem.
Perhaps the most interesting aspect is what this says about the maturity of AI infrastructure. Early AI integrations often focused on connecting models with external tools. The next stage is making those connections scalable, observable, and maintainable under production workloads. A stateless protocol, standardized extensions, stronger authorization, and predictable evolution are not flashy features, but they solve the kinds of engineering problems that determine whether an ecosystem can grow beyond prototypes.
The 2026-07-28 release is therefore less about adding another capability and more about making the entire Model Context Protocol ecosystem easier to operate at scale. For developers building MCP servers or AI-powered applications, the migration requires some work because the release introduces breaking changes. The payoff is a protocol that better matches modern cloud infrastructure and enterprise deployment practices, reducing operational complexity while providing a clearer foundation for future extensions.