New MCP roadmap: agent identity replaces pasted API keys

The Model Context Protocol roadmap landed on August 22, and the most useful phrase in it is about credentials. The maintainers want agent authentication to move away from what they call "pasted API keys and long-lived tokens." If you have wired an MCP server to your CRM, your warehouse, or your billing system, that sentence is a preview of your next two quarters of integration work.

What actually happened

The MCP project published an updated roadmap covering the next specification release and beyond. It is organized into five priority areas, and the governance detail matters as much as the technical one: Specification Enhancement Proposals inside those areas get expedited review, while proposals outside them are not rejected but wait for scarce maintainer time.

The five areas:

Agentic messaging primitives. Maturing the Tasks extension (SEP-2663) toward inclusion in the spec, so a server can accept long-running work, push results back, and let a caller steer a job mid-flight instead of blocking on request-response.

HTTP-native transport unification. Extending the stateless core shipped in the 2026-07-28 specification to more deployment modes, including local servers speaking Streamable HTTP over stdio. One transport story for local and remote.

Agent identity and enterprise security. Finalizing Demonstrating Proof of Possession (DPoP) and driving adoption, plus Workload Identity Federation as "an opinionated path for agent identity and delegation" — the foundation for the ID-JAG grant behind Enterprise-Managed Authorization.

Improved primitives. Standardized tool result handling, and progressive discovery, where a server "can offer a small entry point and reveal more of its catalog as the conversation narrows." The stated reason: tool selection gets worse as the list grows.

SDK developer experience, including conformance testing across platforms.

Why the MCP roadmap matters for your business

Your .env file is on a clock. Right now most MCP deployments are a vendor token in a dotfile, copied onto whatever box runs the agent. DPoP binds a token to a key the client must prove it holds, so a leaked token by itself stops being enough. Workload Identity Federation means the agent gets a short-lived credential derived from an identity your directory already issues. Start by inventorying every long-lived key your agents currently carry — that list is the migration.

Progressive discovery is a cost line, not a feature. Every tool definition you expose is tokens billed on every turn, and accuracy drops as the menu grows. Teams that jammed forty tools into one server are paying twice. Design servers now with a narrow entry point and drill-down, and you get the token savings before the spec makes it official.

Tasks changes what you can hand off. Once long-running work is a first-class primitive instead of a timeout you engineer around, overnight reconciliation, catalog rebuilds, and multi-step fulfillment become things an agent can own with a progress bar you can watch.

Sources: Model Context Protocol roadmap, MCP 2026-07-28 specification.