MCP Dropped Memory from Its Core to Run More Easily Everywhere

When an AI agent needs to open a file in Figma, read a Zoom meeting or call a tool inside a company, one protocol is trying to make the connection standard rather than a fresh invention each time. That protocol, MCP, has now changed its own core.

The fifth version of the specification, MCP 2026-07-28, moves the core from a two-way connection that retains session state to a stateless request-response model. It sounds like a small engineering decision, but it removes one of the heaviest obstacles to deploying MCP servers across serverless infrastructure, edge platforms and enterprise networks.

The server no longer has to remember the conversation

Under the old design, the server and client had to maintain a shared connection and session. That suits some applications, but makes scaling difficult. Where is session state stored? What happens when requests move among multiple instances? And how does a short-lived function operate on an edge platform if the protocol assumes a persistent link?

The new core treats every request as a self-contained unit carrying the context it needs. An MCP server therefore becomes more like a conventional HTTP workload that can be balanced, started and stopped without special session-management engineering. Anthropic says MCP SDK packages have surpassed 400 million monthly downloads, four times the figure a year ago, so scaling is no longer a detail confined to small laboratories.

But “stateless” does not mean agents have lost the ability to complete long tasks. The specification separates simplicity from complexity. The core remains lightweight, while heavier capabilities move into official, versioned extensions.

Apps and tasks leave the grey area

MCP Apps and Tasks are now part of a standard extension framework. Apps allow a server to present an interactive interface inside a conversation, so users can see what a tool is doing and work with it without moving to another tab. Tasks provide a route for work that takes longer than a single request cycle.

The point is that the protocol does not have to expand its core whenever a new capability appears. Extensions can evolve through declared versions while the basic connection model continues to work in the simplest environments.

Authentication moves closer to enterprise reality

The new release tightens authorisation to align with the OAuth 2.0 and OpenID Connect implementations actually used by enterprises. This enables cleaner connections to identity providers such as Entra and Okta instead of requiring a workaround for every connector.

At the same time, Claude is expanding the layer around the protocol. According to Anthropic, its connector directory contains more than 950 MCP servers. The company is also adding centralised authorisation management that lets administrators configure a connector for an entire organisation through identity groups, developer dashboards showing adoption, errors and latency, and MCP tunnels that connect Claude to servers inside private networks without opening inbound ports to the world.

Why does this matter to an Arab developer?

If you are building a connector for a financial, education or government service in the region, the shift reduces the infrastructure you need to manage before the product reaches a single user. Larger parts can be deployed on familiar platforms, with enterprise identity connected instead of a parallel permissions system invented from scratch.

But migration is not simply a change of version number. A server that relied on implicit connection state will need to reconsider where context is stored, and a client must negotiate clearly which extensions it supports. Support for the specification across Claude products is also still rolling out gradually. It is not a switch that appeared everywhere at the same moment.

The most important quality of the new MCP is that it does not try to make every task simple. It makes the foundation simple, then gives complexity an official place where it can be seen and managed. Standards sometimes scale by adding more layers. This time, MCP scales because it recognised what had to leave the core.