The Model Context Protocol's newest revision is now the thing to plan around if you ship retrieval servers, and the headline is architectural rather than cosmetic. The release candidate framed it plainly: a stateless specification , and the tradeoff its maintainers describe is bigger payloads in return for statelessness — "very compressible and very well understood," and still small next to an ordinary HTTP request .
For a knowledge server, that tradeoff lands in a specific place. Session state is where most teams quietly parked the things that make multi-turn retrieval work: rewritten queries, pagination cursors, the reranker's view of what the agent already saw, per-conversation ACL resolution caches. Under Streamable HTTP — introduced in the 2025-03-26 revision to replace HTTP+SSE, using a single endpoint with optional server-to-client SSE — that state had a natural home. Now you either round-trip it in the request or persist it yourself, keyed by something you control. Horizontal scaling gets much easier; free continuity goes away. If your chunk IDs or cursor tokens were only meaningful inside a live session, they need to become self-describing.
The auth changes matter more than the transport
Permission-aware retrieval is the whole game in enterprise knowledge, and this is where the spec did the most work. Six SEPs harden the authorization specification to align more closely with how OAuth 2.0 and OpenID Connect are actually deployed . Concretely, clients now set application_type during Dynamic Client Registration so authorization servers stop rejecting localhost redirects for desktop and CLI apps (SEP-837), and DCR itself is formally deprecated in favor of CIMD — it keeps working for backward compatibility but will be removed in a future version . If you built onboarding around DCR, that's a migration with a clock on it, not an optional cleanup.
This collides usefully with where enterprise search is already heading. Real-time federated connectors query the source system at search time and return current results, running on the individual user's credentials rather than a shared service account — and that pattern is being forced, not just chosen: Slack's May 2025 API terms restricted bulk export and persistent third-party indexing . Query-time federation with per-user identity is exactly the workload the new auth SEPs are trying to make survivable.
Two more items to put on the sprint board. Anyone who shipped against the 2025-11-25 experimental Tasks API needs to migrate to the new lifecycle — that's your reindex jobs and long-running deep-search calls. And out-of-band server logging, where a server could push informational log messages to a client, is gone or narrowed . Retrieval servers leaned on that channel to explain why a query returned nothing. Replace it with your own telemetry before you find out in production.