MCP finally fixes its fatal flaw—and breaks everything you built

MCP just underwent its most significant change since launch: a shift from bidirectional stateful protocol to a stateless request-response model. This means MCP servers can now deploy on serverless infrastructure like Lambda or Cloudflare Workers, with zero cost when idle and trivial scalability. The old requirement of maintaining persistent, dedicated connections between every client and

server—even when tools weren't being called—was a massive overhead that made MCP impractical for real agent workflows. The practical impact is immediate. Previously, running a single agent prompt in tools like Codeex could spawn five sub-agents, each requiring five stateful MCP connections—30 bound processes on one machine. This was the hidden CPU killer behind 'syspolicyd'

on macOS. Now each agent simply hits an endpoint and gets JSON back. No session IDs, no sticky routing, no state management. It's what MCP should have been from the start. However, this introduces a classic standards problem: backward incompatibility. Old clients expect stateful connections; new servers are stateless. The same 'MCP' label now means

two different things, creating fragmentation. The upside is that in an era where models can rewrite integrations in hours, tools that fail to update quickly reveal their incompetence. This becomes a useful filter for which AI tooling companies actually ship. The tradeoff is real: a cleaner spec that breaks interoperability risks undermining the very standardization