interests /

MODEL CONTEXT PROTOCOL WARM

a protocol rewriting itself around what teams could not deploy

4 entries · 2026-06-19 → 2026-08-23 · last entry 44d ago

MCP spent 2026 being rewritten by what teams could not actually deploy. The enterprise auth extension admitted per-user consent was blocking adoption; the July spec dropped the stateful session; the roadmap made that the direction rather than a one-off. Read in order, it is a protocol discovering that its hard problems were operational rather than conceptual.

POST 2026-06-19 [ x ]

MCP Auth Grows Up: One Login for Every Agent Tool

The first sign the protocol was being reshaped by deployment reality rather than by design. The Enterprise-Managed Authorization extension was less about OAuth plumbing than an admission that per-user consent had been quietly strangling MCP adoption inside companies.

The interesting part of the Enterprise-Managed Authorization extension isn’t the OAuth plumbing — it’s the admission that per-user consent was quietly strangling MCP adoption inside companies.

read the post →

later →

POST 2026-07-29 [ x ]

MCP Goes Stateless, and That's the Real Release

The 2026-07-28 spec, and the turn the whole story runs on. MRTR, header-based routing and the extensions framework all read as headline features, but every one of them is downstream of a single decision: the core stopped being a stateful bidirectional session.

The headline features in the 2026-07-28 MCP spec — MRTR, header-based routing, an extensions framework — are downstream of one decision: the protocol core is now stateless. MCP went from a bidirectional stateful session to plain request/response, and every request is self-describing. That’s the change teams running MCP servers in production actually asked for.

read the post →

← earlierlater →

POST 2026-08-05 [ x ]

Stateless MCP Is the Version That Fits Production

The same spec change a week later, read from operations rather than design. A tool call becoming one HTTP request sounds like protocol trivia until you have tried to run a stateful MCP server behind a load balancer — which is where the change stops being cosmetic.

The headline change in the 2026-07-28 MCP spec is that a tool call is now a single HTTP request instead of an initialize-then-call handshake. That sounds like protocol trivia until you’ve tried to run a stateful MCP server behind a load balancer. Simon Willison’s write-up frames it as a renewed personal interest; from an ops seat it reads as MCP finally shedding the one property that made it painful to operate.

read the post →

← earlierlater →

POST 2026-08-23 [ x ]

What the New MCP Roadmap Means If You Ship Agents

The roadmap confirms the stateless core as the direction, but the five priority areas are the more useful signal. They show where the protocol now believes its hard problems live, and it is no longer connection management.

The headline of the new MCP roadmap is the move toward a stateless core, but the more telling signal is what the five priority areas say about where the protocol now thinks the hard problems live. They’ve shifted from “how do I connect a tool” to “how do I operate agents I can trust in production.”

read the post →

← earlier

Select an entry to read it here.