AI’s Essential Protocol is Growing More User-Friendly

The Model Context Protocol (MCP) is a crucial element for AI interoperability, offering a secure method for AI models to access external data sources and services. It functions as the bridge that allows a chatbot to connect with your calendar, database, or internal tools, removing the need for engineers to develop custom connections for each integration. Next week, this protocol will receive a significant update that, while it may go unnoticed by end users, could greatly impact the evolution of the ecosystem.

The official specification for the new version has been publicly available since May, but we got a particularly clear summary of the updates on Monday morning from the team at Arcade—a two-year-old startup focused on enabling AI agents to perform efficiently within real businesses, allowing secure connections and interactions with tools like Gmail, Slack, and Salesforce.

Arcade raised $60 million in funding in June, grounded in the belief that most AI agents falter not because of weak underlying models, but due to insufficient supporting infrastructure—a challenge this update strives to address. Essentially, MCP is overhauling the management of session IDs—small tokens that servers use to remember that “this is the same conversation as five seconds ago”—to ease operations at a larger scale.

As Nate Barbettini, founder of Arcade, explains:

[In the current system] When an MCP client like Claude first connects to a server, it sends a “hello”: I’m Claude, here’s my version, here are my capabilities. The server replies with its capabilities and issues a session ID… Then, the client includes that session ID with every request to signify it’s the same conversation. Occasionally, the ID expires, leading the client to notice, request a new one, and carry on…

Imagine a real-world deployment. You’re managing a server for millions of users, behind a load balancer that directs each request to the available servers, sometimes across various regions. Each of those machines must keep track of a session ID issued by another machine. While it’s not impossible, it adds significant complexity, interfering with the load balancer instead of working in tandem with it.

In short, the current setup depends on a single server to remember user sessions, whereas, in practice, businesses distribute traffic across multiple servers that don’t communicate by default. This results in MCP servers encountering additional challenges in tracking user identities, posing considerable difficulties for those managing MCP servers at scale and partly explaining the lack of extensive first-party MCP integrations despite the current enthusiasm surrounding agentic AI.

With the new system, the protocol will embrace a more adaptable, “stateless” approach to session IDs on the server side, similar to how most conventional websites function, which should streamline maintenance and potentially lower operational costs at scale.

Although this may sound technical, it underscores an important reality: not all facets of AI development are progressing at breakneck speeds. While model training is advancing quickly, much of the technical infrastructure supporting these models still relies on the slower consensus-building process typical of standards organizations. Progress is indeed being made; it simply occurs at a more measured pace!