The 2026-07-28 Model Context Protocol specification is final, and over the past couple of weeks the migration guidance has arrived from the people who host most of these servers. Anthropic published notes on bringing the spec to Claude, Google's developer blog framed it as scaling agent infrastructure, and Microsoft walked through what it changes for running MCP servers on App Service. Trade coverage from InfoWorld and The Register landed on the same summary: MCP is dropping its stateful past to make scaling simpler. A security analysis of the stateless design is also circulating. That mix — vendor migration docs plus a security writeup, all within weeks of the spec date — is the tell that this has moved from spec discussion into planned work.
Why this hits retrieval servers hardest
Most MCP servers in a knowledge stack are thin adapters over something else: an OpenSearch cluster, a vector index, a Confluence or Drive connector, a SQL warehouse. The thin ones are fine. The problem is that a stateful transport quietly invited teams to park things in the session — resolved auth and tenant context, pagination cursors mid-way through a long result set, warm connection pools, a reranker cache keyed to the conversation, incremental filter state built up across tool calls.
A stateless core means every one of those has to move: into the request itself, into an external store, or out of the design. The upside is real and boring in the good way. No sticky routing, no session affinity in the load balancer, no drain-before-deploy dance, horizontal scale that actually works, and a much smaller blast radius when a pod dies mid-query.
The permission problem gets sharper
The expensive one is permission-aware retrieval. ACL-filtered search depends on knowing who the caller is and which groups they belong to, and group resolution against an IdP or a directory is not something you want to do on every tool call. Under a session model you resolved it once and held it. Stateless, you need either a short-lived token that carries entitlements with it, or a cache tier in front of your identity source with an invalidation story you can defend when someone leaves a group. Neither is hard; both are work, and the second one is where stale-permission bugs live.
What to do this week
Inventory what your MCP servers keep between calls. If the answer is "nothing," you get the scaling benefits for free. If it's cursors and cached identity, scope that now rather than discovering it when a host you depend on stops honoring long-lived sessions.
Adjacent and worth watching: connector-layer consolidation continues — Nexla announced it passed 1,000 enterprise connectors, pitched explicitly at governance and context for agents rather than raw pipe count. Governance is where connector plays are competing now, not coverage.