The Model Context Protocol (MCP), the open standard that enables AI models to connect to external applications and data sources, is being updated. The new protocol version replaces the current session-based approach with a stateless architecture. This should make it easier to roll out MCP at scale across enterprises and cloud environments.
MCP has become a key building block for agentic AI. The protocol enables AI assistants and agents to securely communicate with email, calendars, databases, CRM systems, and internal business applications, among others. As a result, developers no longer need to build a separate interface for each integration.
The first versions of MCP were primarily designed for local applications. In this setup, a client, such as an AI assistant, first establishes a session with a server. During this initialization, both parties exchange information about their capabilities, and the server then provides a session ID. The client uses that same session ID for every subsequent request.
This approach works well when the client and server run on a single computer or in a small environment. In large-scale cloud environments, however, it introduces additional complexity. There, requests are routed via load balancers to different servers, all of which must have access to the same session data.
No more session architecture
The new MCP version moves away from that session architecture. From now on, every request contains all the information a server needs to process it. As a result, the server no longer needs to maintain session state between consecutive requests.
This change aligns with how modern web services have been operating for years. Since servers no longer need to synchronize session data, it becomes easier to distribute traffic across multiple systems and support large numbers of concurrent users.
According to TechCrunch, this limited scalability has been one of the reasons many software vendors have been reluctant to offer official MCP servers for production environments. The protocol change is intended to lower that barrier and accelerate the rollout of enterprise implementations.
The transition is not without consequences. Existing clients and servers are built around the current initialization procedure using session IDs and, therefore, cannot communicate with the new protocol version without modifications.
MCP does support version negotiation, allowing implementations to temporarily offer multiple protocol versions side by side. However, developers of clients, servers, and gateways will need to adapt their software to fully support the new architecture.
Less prone to failures
In addition to improved scalability, the new approach is also expected to increase reliability. Problems with expired sessions and disconnected connections will largely disappear once servers no longer need to maintain session state. Implementing MCP servers will also become simpler, as less logic is required for session management.
This change demonstrates that AI development is not solely focused on increasingly powerful language models. The underlying infrastructure is also evolving rapidly. As new AI models emerge rapidly, protocols such as MCP are becoming increasingly important for enabling AI systems to interact securely and in a standardized manner with business applications. As a result, MCP is evolving from a tool for local AI assistants into a protocol suitable for large-scale enterprise environments.