On May 22, 2026, the MCP working groups announced the MCP Release Candidate (RC) — the biggest single revision to the protocol since its original launch. The final release is scheduled for July 28, 2026.
| Change | What It Means | Impact |
|---|---|---|
| Stateless Core | Eliminates sticky sessions and session IDs from the core protocol. MCP servers can now run behind standard round-robin load balancers using plain HTTP. | 🔴 Biggest architectural change — simplifies deployment at scale dramatically |
| Extensions Framework | New capabilities are negotiated as extensions rather than being baked into the core specification. | 🟡 Enables faster iteration without breaking the core protocol |
| Tasks Extension | Formal support for long-running asynchronous operations (evolved from SEP-1686 Tasks Primitive). | 🟢 Critical for agent workflows that span minutes or hours |
| Enhanced Authorization | Aligns MCP auth with modern OAuth 2.1 and OpenID Connect standards. | 🟢 Enterprise-ready SSO and identity federation |
| Formal Deprecation Policy | Establishes long-term stability guarantees with defined deprecation timelines. | 🟢 Confidence for production deployments |
Before the RC, MCP servers were inherently stateful — each client-server pair maintained a session, requiring sticky routing in load balancers. This made horizontal scaling painful:
// BEFORE (stateful — requires sticky sessions): Client A ──▶ Load Balancer ──▶ Server Instance #3 (pinned) Client B ──▶ Load Balancer ──▶ Server Instance #1 (pinned) // AFTER RC (stateless — standard round-robin): Client A ──▶ Load Balancer ──▶ Any Server Instance Client B ──▶ Load Balancer ──▶ Any Server Instance
With the stateless core, MCP servers are now plain HTTP services that can be deployed, scaled, and load-balanced with existing infrastructure — no special session affinity required.