Model Context Protocol revision drops sessions for cloud-scale AI apps

MCP maintainers plan to finalize the next Model Context Protocol revision on July 28, with a shift to stateless requests that changes how AI applications connect to external systems at scale. The update matters because MCP, originally released by Anthropic in 2024, has moved from local developer demos into cloud-hosted enterprise deployments, where its session-based design created operational friction.

David Soria Parra, a member of Anthropic’s technical staff, said in a livestream Tuesday that the 2026-07-28 revision includes the protocol’s most substantial specification changes since authorization was added. He said teams that have kept their SDKs current should have a manageable upgrade path, while developers that built their own MCP implementations should expect more work because the underlying transfer mechanism has been rebuilt.

Anthropic donated MCP in 2025 to the Agentic AI Foundation, a Linux Foundation fund co-founded by Anthropic, Block and OpenAI. By then, according to the maintainers, the SDK was seeing more than 97 million downloads per month and at least 10,000 MCP servers had been deployed.

What is changing in the Model Context Protocol revision?

The main change is the removal of protocol-level session tracking. In the new design, each request can be processed independently, with details such as protocol version, client identity and capabilities carried in the request’s _meta parameter rather than stored in a server-side protocol session.

Stateless design is common in cloud services because requests can be spread across load balancers and server fleets without requiring infrastructure to remember which client belongs to which session. For MCP, that is a material shift from its original local-tooling roots, where servers exposed resources, tools and prompts to LLM-connected applications through clients using JSON-RPC 2.0.

Parra said the first design left important session behavior underspecified, which made implementation and routing harder as deployments became more complex. He also said the new wire protocol is more complicated than the old one, even if the result should be easier to operate.

Routing is also being reworked. Earlier MCP messages put the operation inside JSON-RPC HTTP POST payloads, which meant routing systems had to inspect the JSON body or depend on session state. The new revision mirrors routing information in HTTP headers, allowing network infrastructure to route requests without tracking MCP sessions.

Which MCP features are being deprecated?

Several features from the original protocol are being deprecated. Sampling, which let an MCP server ask a client’s model to generate a completion, was rarely used and difficult to define cleanly, according to Parra. Roots, a feature for telling servers which filesystem locations matter to a session, is also being deprecated. Logging is being scaled back as well, with developers pointed toward stderr, stdio or OpenTelemetry.

The maintainers are adding a formal extension model to move narrower or experimental capabilities outside the core specification. Den Delimarsky, also on Anthropic’s technical staff, said extensions are intended to test protocol ideas before they are considered for the core.

One official extension is MCP Apps, which lets MCP present an interactive JavaScript application rather than only text or an image. Tasks, used for long-running operations, is moving from the core protocol into an extension. Parra said users will also be able to build their own extensions under the specification’s guidance.

Stacklok, which sells an MCP platform, said in an Enterprise Readiness Guide that deprecated features will keep working for at least 12 months. The company also warned that this does not guarantee broad compatibility: servers using the 2026-07-28 revision may not interoperate with older clients unless one side supports fallback or translation.

Stacklok CEO Craig McLuckie, a Kubernetes co-creator, told The Register that MCP’s original statefulness reflected its start as a local coding-tool protocol. He said broader enterprise adoption is pushing MCP toward a role as a control point for access to production systems, including policy enforcement by cybersecurity teams.

This story draws on original reporting from The Register.