The Model Context Protocol maintainers published an updated roadmap on August 22, and the item that should concern anyone building knowledge connectors is buried in the fourth priority area: progressive tool discovery. It's an admission that tool catalogs have become a retrieval problem, and the protocol currently has no ranking layer.
What already shipped
The 2026-07-28 specification, published July 28, is the largest revision since MCP launched. The core is now stateless. The `initialize`/`initialized` handshake (SEP-2575) and the `Mcp-Session-Id` header (SEP-2567) were removed outright — not deprecated. Protocol version, client info, and client capabilities now ride inline in a `_meta` field on every request, so any request can be served by any instance behind a plain round-robin load balancer. Google, which co-founded the MCP Transports Working Group with Hugging Face and others, notes that running remote servers previously meant shared Redis session stores or gateway-level packet inspection. That cost is gone; serverless and edge deployment are now viable.
Supporting changes matter as much as the headline. `server/discover` lets a client learn supported versions and capabilities up front, list results became cacheable (SEP-2549), Tasks became a formal extension (SEP-2663), and Multi Round-Trip Requests (SEP-2322) replaced server-initiated requests so elicitation still works without a session. Authorization picked up issuer validation, issuer-bound client credentials, and Client ID Metadata Documents as the preferred client registration path.
Why the catalog is the next bottleneck
Cacheable list results reduce the wire cost of enumerating tools. They don't reduce the context cost. If you expose a document store, a ticket system, and three internal APIs through MCP, the agent pays for every tool description in every turn before the user has asked anything. Developers Digest, which runs an MCP-based skills platform, named this directly: tool catalogs that consume context ahead of any query. The Server Card Working Group's `.well-known` metadata conventions point the same way — letting a client reason about a server without connecting to it.
The practical read: treat your tool surface the way you treat a corpus. Rank, filter, and load descriptions on demand. If you're shipping fifty tools behind one endpoint today, you're already eating the tax, and the protocol fix is a roadmap item, not a spec.
Identity is the other half
The roadmap's agent identity work — DPoP, Workload Identity Federation, token exchange — matters most to anyone doing permission-trimmed enterprise retrieval. Stateless requests mean authorization decisions happen per-request, per-caller, with no session to hang a user context on. Enterprise-Managed Authorization is stable as an extension, and Anthropic's MCP tunnels research preview reaches private-network servers without public endpoints or inbound firewall rules.
Migration is not free. Old clients only keep working if you run both paths, and Roots, Sampling, and Logging survive until at least July 2027. Plan the dual-serve window now.