MCP Just Went Stateless: What the July 28 Spec Rewrite Breaks and Fixes
The Model Context Protocol's July 28, 2026 spec update is not a routine version bump. It introduces a stateless protocol core, multi round-trip requests, header-based routing, and cacheable list results, alongside a formal extensions framework for future capability additions. If you have production MCP servers, this changes real behavior, not just documentation.
The change with the most practical bite: Dynamic Client Registration is formally deprecated in favor of CIMD, and the spec itself acknowledges some of these changes are not backward compatible. That's a meaningful statement for a protocol this widely deployed. Tier 1 MCP SDKs see close to half a billion downloads a month, and the TypeScript and Python SDKs have each individually crossed a billion total downloads. A lot of infrastructure is now sitting on top of a protocol that just changed its foundational handshake model.
Why stateless matters
Moving the protocol core to stateless operation is a scalability decision as much as a design one: stateful connections are harder to load-balance, harder to recover cleanly after a crash, and harder to reason about at the scale MCP has actually reached. Cacheable list results and header-based routing are the practical payoff, servers can now be deployed behind standard HTTP infrastructure patterns instead of needing session-aware routing built specifically for MCP's old connection model.
What to check before you upgrade
- Does your integration rely on Dynamic Client Registration anywhere in its auth flow? That path is deprecated now, not just discouraged.
- Are you assuming session state persists across requests anywhere in your server implementation? The stateless core removes that guarantee.
- Have you tested against the new spec in a staging environment before assuming your existing integration is unaffected, given the spec explicitly flags non-backward-compatible changes?
MCP's growth curve means a spec change like this touches a genuinely large surface area of production AI tooling. Worth treating this upgrade with the same care you'd give any other foundational protocol change, not as a routine dependency bump.