Why the old MCP was a resource hog

The original MCP spec required a dedicated, stateful connection between client and server for the entire duration of a session — even if the tool was never called. That meant every MCP server you configured had to stay bound to your agent the whole time, whether you used it or not.

In practice this created absurd overhead. A single Codex run spinning up several sub-agents could result in dozens of simultaneous stateful connections on both the client and server side, hammering local CPU (visible on Mac as heavy syspolicyd activity) and making cloud hosting of rarely-used MCP servers not worth the infrastructure cost.

Understand the old bottleneck: every MCP server needed its own live connection regardless of whether it was actually used.
Recognize the symptom: multiple sub-agents multiplied connections fast, which is why many teams just ran MCP servers locally instead of in the cloud.

What the new stateless spec actually changes

Anthropic's newest MCP release (versioned by date, not a clean name) moves the protocol from a bidirectional, stateful model to a simple request-response model. A server can now be a plain HTTP endpoint — deployable on serverless or edge infrastructure — instead of a long-lived process tied to a load balancer that has to route requests back to the same instance.

Simon Wilson, an early skeptic-turned-tester, built several MCP tools in the days after the release specifically because implementation got so much simpler. Where the old spec needed two HTTP requests (one to open a session, one to call a tool) plus session-ID routing, the new one needs a single request with no session state to track.

Deploy anywhere: stateless servers can run on Lambda, Cloudflare Workers, or similar, and cost nothing when idle.
Skip the load balancer complexity: no more sticky sessions or matching session IDs to specific backend instances.
Expect smaller, cleaner MCP servers: several previously-abandoned MCP projects became shippable once the stateless model removed the infrastructure burden.

The compatibility problem nobody can avoid

The new spec is not backward compatible. Old clients expect a bound, stateful connection; the new servers don't provide one. That means a tool advertising "MCP support" could mean the old heavyweight standard or the new lightweight one, and there's no way to tell without checking versions.

This creates real friction during the transition: a server built on the new spec may simply fail to connect to an agent harness — Claude Code, Codex, Cursor, or others — that hasn't updated its MCP client implementation yet.