Model Context Protocol (MCP) is one of the fundamental building blocks of AI interoperability, providing AI models with a secure way to access external data sources and services. This is plumbing that allows chatbots to access calendars, databases, or internal tools, rather than having engineers build custom pipes for each connection. Next week, the protocol is set to receive a major update that may not be noticeable to end users but could lead to major changes in the evolution of the ecosystem.
Official specifications for the new version have been available since May, but on Monday morning we received an unusually clear explanation of the changes from Arcade officials. Arcade is a two-year-old startup that has built its entire business around putting AI agents to work within real enterprises, securely connected to tools like Gmail, Slack, and Salesforce.
Arcade raised $60 million in June based on the idea that most AI agents fail not because the underlying model is weak, but because the surrounding infrastructure isn’t ready yet. And that’s what we’re trying to solve with this update. Essentially, MCP changes the way session IDs (the little tokens that servers use to remind themselves, “Oh, this is the same conversation I had five seconds ago”) make it easier for servers to work in larger environments.
Nate Barbettini, founder of Arcade, said:
When an MCP client like Claude (in the current system) connects to the server for the first time, it sends a “Hello” message. I’m Claude. This is my version and here is my functionality. The server responds with its own capabilities and returns a session ID. From then on, the client sends that session ID in every request, so the server knows it’s the same conversation. In some cases, an ID expires and the client must be aware of it and request a new ID to proceed.
Please imagine the actual development. We run servers for millions of users behind a load balancer. The overall role of the load balancer is to route each request to an available server within the farm (possibly in another region). All of these machines must know the session ID distributed by the other machines. It’s not impossible, but it’s a serious problem, and you’ll be competing with your load balancer instead of working with it.
In other words, today’s MCP servers have to do extra work just to keep track of who is who, because current setups assume one server remembers users, but real-world enterprises spread traffic across dozens of servers that don’t communicate with each other by default. This is a huge headache for those running MCP servers at scale, and is one of the reasons why, despite all the hype around agent AI this year, we haven’t seen more companies ship first-party MCP integrations at scale.
In the new system, the protocol will take a more relaxed “stateless” approach to server-side session IDs, similar to how most regular websites already work. This should make the whole system much easier to maintain, and theoretically cheaper to run at scale.
This is all fairly technical, but it’s an important reminder that not all parts of AI development are moving at breakneck speed. While model training is moving forward, much of the technical infrastructure needed for models is still subject to the slow logrolling of standards body agreements. It’s actually happening. It’s just a little late!
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
