Model Context Protocol prepares to break with its stateful past - The Register

What the report says

The Register reported that maintainers of the Model Context Protocol, or MCP, are preparing a major revision scheduled for finalization on July 28, 2026. MCP, introduced by Anthropic in 2024 as an open standard for linking AI applications with external systems, is being reworked after its use expanded from local developer demos to cloud-hosted, multi-client enterprise deployments. Anthropic technical staff member David Soria Parra described the update in a livestream as the protocol’s largest specification change since authorization was added.

The central shift is the removal of protocol-level sessions. Under the revised design, requests can be handled independently, with details such as protocol version, client identity and capabilities carried in each request’s metadata. The Register said the change is intended to make MCP more like stateless web infrastructure and easier to operate behind load balancers. Routing information will also be reflected in HTTP headers, reducing the need for networking systems to inspect JSON-RPC bodies or maintain session context.

The update is not fully backward compatible. Parra said users who have kept their SDKs current should face fewer problems, but warned that custom MCP implementations may require substantial rework. Several features are also being deprecated, including sampling, roots and built-in logging, while tasks are moving from the core protocol into an extension. A new extensions model is intended to let maintainers test or support domain-specific capabilities without adding every feature to the base specification.

The Register noted that Anthropic donated MCP in 2025 to the AAIF, a Linux Foundation fund co-founded by Anthropic, Block and OpenAI. Stacklok’s guidance, cited by the publication, says deprecated features should keep working for at least 12 months, but compatibility between old and new clients and servers is not guaranteed without shared protocol support or fallback mechanisms.