Anthropic has delivered the largest revision yet to the Model Context Protocol, or MCP, shifting the protocol’s core from a stateful design to a stateless one. According to the official MCP blog, the MCP 2026-07-28 specification released on July 28 allows MCP servers to run directly in serverless and edge computing environments.
The update marks the most dramatic architectural change since MCP was introduced.
Handshake and session ID removed
Earlier versions of MCP used a bidirectional, stateful protocol. A client first had to complete an initialization handshake with the server and obtain a session ID, listed as Mcp-Session-Id. After that, every request had to return to the same machine.
That setup made horizontal scaling difficult and ruled out serverless deployment. The new specification removes both the initialization handshake and the session ID, replacing them with a request-response model. Each request now carries identity information through the _meta field and HTTP headers, which means any machine can handle it. A standard load balancer in front of the service is enough to scale the system out.
That change is central to MCP’s ability to run in serverless and edge environments. The report also said the specification has already made its way into Amazon Bedrock AgentCore.
MRTR added alongside three official extensions
To preserve interactivity without a persistent connection, the new specification introduces Multi Round-Trip Requests, or MRTR. The mechanism lets tools ask users for confirmation while a task is still in progress, such as prompting with a question like “Do you want to delete this row?”
The update also formally incorporates three official extensions:
- MCP Apps, a server-rendered user interface
- Tasks, for asynchronous and long-running operations
- Enterprise Managed Auth, an authorization model based on identity providers and organization-wide provisioning
The specification also strengthens its authorization framework.
Monthly SDK downloads top 400 million
The scale of the revision also reflects how widely MCP is now being used. Official figures show monthly SDK downloads have recently exceeded 400 million, with usage up 4x this year. MCP is described as an industry standard for connecting AI agents with a wide range of applications.
By making the protocol core stateless, MCP servers can now be scaled and operated more like ordinary internet services, laying the groundwork for broader enterprise and cloud deployment.