MCP just got its biggest spec change since launch and it is going stateless

[RadekVítek]

The Agentic AI Foundation, which is a Linux Foundation directed fund, released the 2026-07-28 MCP specification and this is being described as the largest change to the protocol since it originally launched

The headline change is that MCP moves to a stateless request response core, which eliminates the initialize and initialized handshake along with the Mcp-Session-Id header entirely, meaning any server instance can now handle any request without needing sticky session state tied to a specific process

If you have ever had to deal with load balancing MCP servers behind a proxy you know exactly why this matters, session affinity requirements have been a genuine pain point for anyone trying to run these things at any real scale in production

Going stateless also opens the door to much simpler horizontal scaling and easier serverless deployment models, since you no longer need to route a given client's follow up requests back to the exact same server instance that handled the handshake

The tradeoff worth watching is whether stateless design pushes more complexity back onto the client side or onto whatever session management layer developers now have to build themselves on top, protocols rarely eliminate complexity so much as relocate it

The 2026-07-28 Specificationblog.modelcontextprotocol.io

[DeBruyne75]

Finally, session affinity in MCP has been the single most annoying thing about deploying this at scale for the last year

[EventHorizon63]

Genuinely relieved to see this, our infra team has been building increasingly hacky sticky routing workarounds for months

[Jonathan]

Curious how backward compatible this is with existing MCP servers, is this a breaking change everyone has to migrate for or additive

[EdgeNode Joel]

Stateless is the right long term call but I bet a lot of tooling built around the old session model quietly breaks in ways nobody notices until production

[Donna75]

From the spec notes it looks like a breaking change for anything relying on the old handshake, so expect a wave of compatibility issues over the next few months

[Connor97]

The comparison to how HTTP itself evolved feels apt here, statelessness at the protocol layer with state pushed to the application layer is basically the same lesson relearned

[Coder58]

Does this change anything for how tool call authentication works or is that layer completely separate from the session handshake being removed