MCP's Maintainers Want the Standard to Be Boring, on Purpose

MCP, the Model Context Protocol and open standard that lets AI agents call external tools and services, just got a mid-cycle update. The new roadmap reads as consolidation, not expansion.

Three beats shape the next phase. First, transport: maintainers are pushing toward one HTTP-native transport, the same plumbing that runs most web services, so a remote MCP server runs like any other HTTP workload instead of needing bespoke hosting. The work builds on the 2026-07-28 spec release and extends that transport to local servers over stdio.

Second, agent identity. Today, MCP authorization assumes a person in a browser. The roadmap describes the agent identity work as built on existing standards rather than pasted API keys. The specifics — which standards and authorization grants will carry the weight — remain open.

Third, primitives for long-running work: server-initiated events like webhooks and channels, a maturing Tasks extension (SEP-2663), and standardized result contracts with progressive discovery so a server reveals its catalog as a conversation narrows.

These are maintainer priorities, not adoption proof. Standardizing identity for non-human callers is genuinely hard, and the roadmap leaves the specific standards undefined. Unifying on one transport is a simplification story for new servers and a compatibility story for existing ones.

The honest read: MCP is being shaped for deployment, not demos. Boring on purpose.