On the 28th of July the Model Context Protocol, the plumbing that lets AI agents talk to your tools and your data, published its biggest revision since it added authorisation. The headline in one line: MCP is now stateless. No handshake, no session id, any request can hit any server instance. If you run MCP servers, that changes what your infrastructure has to do, and mostly for the better.
What “stateless” means for your wallet
Under the old design, a remote MCP server needed sticky sessions, a shared session store, and deep packet inspection at the gateway to keep track of who was mid-conversation. That is fiddly and expensive to run at any scale. The new revision moves MCP to a plain request-and-response model, so the same server can now sit behind an ordinary round-robin load balancer and run on serverless or edge infrastructure. In practical terms, a class of hosting complexity you were paying for has been removed. Cheaper to run, easier to scale, less to break at 3am.
Two extensions worth knowing: Tasks and MCP Apps
The revision also makes extensions a first-class part of the protocol, so new capabilities can move on their own timeline without dragging the core with them. The first two matter. Tasks, contributed by AWS, is built for long-running work: instead of holding a connection open and praying, a server can answer a tool call with a task handle, and the client drives it forward with tasks/get and tasks/update. That is how you get reliable agents that run for minutes or hours without falling over. MCP Apps lets a server send interactive HTML that the AI client renders inside a sandboxed iframe, with the UI declared ahead of time so the host can prefetch, cache and security-review it. Every button the user clicks goes back through the same audit and consent path as a normal tool call, so you get real widgets in the chat without handing the server a blank cheque.
The catch, and it is a real one
This is not backward compatible. Servers built on the 28 July revision may not work with older clients, and the reverse is also true, so you cannot upgrade one half of the pipe and forget the other. The good news is that the maintainers added a formal deprecation policy for the first time: anything marked deprecated stays working for at least 12 months, so you get warning before the floor moves. Anthropic has already said it is bringing the new revision to Claude, so this is not a paper standard.
What to do about it
If you maintain an MCP server, read the migration notes now rather than in a panic later, and plan the client-and-server upgrade as one job. Your reward is cheaper hosting and proper long-running Tasks. If you only use agent tools, you do not have to lift a finger, but expect two visible changes over the coming months: more interactive panels showing up inside your AI client, and long jobs that stop timing out on you. Boring infrastructure news, real everyday payoff.
Did you know: the load balancer, the humble box that spreads requests across servers, dates to the mid-1990s, when a single web server melting under traffic was a genuinely novel problem to have.