The Model Context Protocol shipped its 2026-07-28 specification, published alongside a release candidate post on the MCP blog the same day. Anthropic's announcement headline names the change directly — a stateless core, arriving in Claude. Obot had flagged the same theme a week earlier when reading the 2026 roadmap: stateless core, extensions, enterprise readiness. If you maintain connectors that expose an internal knowledge base, a vector index, or a document store to an agent, this is the release that affects your deployment topology.
Why statelessness is the load-bearing change
Session-oriented MCP servers were straightforward to write and unpleasant to operate. A server holding per-session state has to pin a client to an instance, which means sticky routing, session affinity at the load balancer, and awkward failure behavior when a pod dies mid-conversation. That is tolerable for a local server reading your filesystem. It is a real cost for an enterprise search connector fronting Confluence, a ticketing system, and three object stores, where you want horizontal scale-out and no meaningful recovery path other than retry.
A stateless core pushes that state to the edges — into the client, into tokens, or into explicitly declared extensions. The practical read for connector authors: your server becomes closer to an ordinary stateless HTTP service, which is the shape your platform team already knows how to run. The cost is that anything you were quietly keeping in session memory — cursors, cached auth context, partially materialized retrieval results, ranking state across a multi-turn refinement — now needs an explicit home. Pagination over large corpora is the case worth auditing first, because "hold the result set and hand out pages" is exactly the pattern statelessness removes.
Authentication was the other pressure point. WorkOS had been writing since June about this spec cycle changing agent authentication; the ecosystem work suggests the same. Microsoft shipped v2.0 of the official MCP C# SDK on release day, and Cloudflare followed on August 6 with its own post on the next generation of MCP. When the runtime vendors move in the same week as the spec, the compatibility window for older servers starts closing.
The adjacent thread worth watching
Agent memory evaluation is consolidating at the same time. Agentry covered a MemoryArena benchmark at the end of July framed around a gap in how agent memory gets evaluated, and Mem0 published a State of AI Agent Memory 2026 report with benchmarks. Vectorize maintains an open agent-memory benchmark on GitHub. None of this has settled into a standard yet, but if you are choosing a memory layer on vendor claims alone, there are now at least three independent yardsticks to check them against.
Migration advice is unglamorous: inventory every MCP server you run, find the session state, and decide whether it moves to the client or becomes an extension.