AI’s most important protocol is getting a little bit easier to use
TechCrunch Russell Brandom
MCP, the protocol letting AI models plug into outside tools and data, is getting an update to how it tracks sessions. It sounds boring but it's the fix that could finally let companies run these integrations at real scale.
Based on reporting by TechCrunch, Russell Brandom — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a decent chance you've never heard of the Model Context Protocol, and that's fine, because it's not built for end users. It's the connective tissue that lets something like Claude reach into your calendar or an internal database without an engineer hand-wiring every single connection. Boring, essential plumbing. And next week, that plumbing is getting a fix that's been a long time coming.
The issue was session IDs. Right now, when an MCP client connects to a server, the server hands back a token so it can recognize "oh, this is the same conversation from a few seconds ago." Fine for a single machine. But Nate Barbettini at Arcade laid out on Monday why this falls apart the moment you're running millions of users behind a load balancer: that load balancer's whole job is to spray requests across dozens of servers in different regions, and now every one of those servers needs to somehow know about a session ID some other machine issued. It works against the load balancer instead of with it, and it's been a genuine pain for anyone trying to deploy MCP at any real scale.
The new spec, public since May but only now getting a clear public explanation, ditches that assumption. Servers go stateless on session tracking, which is basically how most regular websites have operated for years. No single machine needs to remember you specifically. That should make large MCP deployments cheaper to run and far less annoying to maintain, and it may be part of why we haven't seen more big first-party MCP integrations this year despite all the noise around agentic AI.
None of this will show up in a changelog anyone outside infrastructure teams reads. But it's a useful data point on how AI actually develops. Model training moves at a sprint. The standards and plumbing underneath it move at the pace of committee consensus, which is to say: slowly, with public comment periods, and real engineers arguing about session tokens. It's still happening. It's just happening on a very different clock than the one everyone's used to watching.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous work that actually determines whether agentic AI becomes real infrastructure or stays a demo reel, and I'd bet on protocol fixes like this mattering more long-term than the next model release. Open, boring standards win when they let anyone build on top without asking permission — that's the whole architecture of the useful internet, and MCP finally acting like ordinary web infrastructure is a good sign, not a letdown.
Read more about this at: TechCrunch