What's this about?
Model Context Protocol — the now-standard way for AI models to talk to external tools and data — shipped a major new version dated 2026-07-28. The headline change: it became "stateless." Previously a client connected, got a session ID, and had to keep talking to that exact same server instance. Now every request stands alone, so MCP servers can be deployed like any ordinary web service — behind a load balancer, on serverless, at the edge — instead of pinning each user to one machine.
What exactly is "stateless vs. stateful"?
"State" is memory the server keeps about your ongoing conversation. A *stateful* server is like a phone call: you have to stay connected to the same operator who remembers your context. A *stateless* server is like sending self-contained postcards: each request carries everything needed, so any available server can handle it, and you can add or remove servers freely to handle load. The tradeoff is that if a tool genuinely needs to remember something between calls, that memory now lives in the tool's own data (a "handle" the model passes along), not in the protocol.
What's the community saying?
- Relief from operators: someone running an MCP gateway/registry (Glama) said a huge share of their bugs came from needing to persist server state — this change removes that whole class of problems.
- "This is just correct HTTP": several argued remembering state was always the client's job, not the server's, and that stateless request/response is exactly why plain HTTP scaled so well.
- The skeptics: do we even need MCP for most cases? One argued a SKILL.md file with linked docs and plain curl commands beats it for many workflows.
- Migration questions: can I keep running MCP 1.x for now, and is it possible to adopt 2.x piece by piece starting with the stateless core? (Maintainers were in-thread answering.)