MCP 2026-07-28 opts for stateless communication

An AI agent needs a tool. Its MCP client asks the server which tools are available, reads their names and parameters, and then calls the appropriate tool. The version 2026-07-28 of the Model Context Protocol (MCP), published on 28 July, a standardized interface between AI applications and data sources, now fundamentally changes the process under the hood.

Previously, the client and server could negotiate sessions at the beginning. The new version switches communication entirely to a stateless model: there are no longer any sessions for the server to manage. Technically, requests are transmitted over HTTP and structured using JSON-RPC 2.0, a method for exchanging commands and results.

For server operators, this is primarily an operational change. They no longer need to track sessions, and the load can be distributed more easily across multiple components in the background. This is intended to simplify the architecture. At the same time, the version contains a series of Breaking Changes that developers will have to take into account going forward.

The basic principle remains straightforward: an MCP server publishes tools, resources as context-only data, and prompts as ready-made templates. The client retrieves the list, the model selects a tool when needed, and the result flows back into the dialog. The protocol's developers compare MCP to a USB port for AI applications.

And then? For teams connecting AI agents to multiple data sources, the practical benefit lies in operations: less session logic on the servers and simpler request distribution. But the migration is not an invisible detail. Anyone already using MCP must bring the client and server over to the new stateless communication model; heise online also cites a separate server in fourteen lines as an example.