The protocol connecting AI to your tools just got rebuilt

On July 28 the Model Context Protocol shipped its fifth spec revision, and the people who run this plumbing for a living have been posting about it all week. @ClaudeDevs called it "the largest update to the protocol since launch" and led with the headline change: MCP is now stateless.

If you have never heard of MCP, you have almost certainly used it. It is the standard that lets an AI assistant reach into something else you own: your calendar, your CRM, your file storage, your job scheduler. When you connect Claude or ChatGPT to a tool and it starts reading and writing real records instead of guessing, that connection is usually an MCP server on the other end. Anthropic published the protocol in late 2024 and then handed it to a foundation. The official SDKs now pull roughly half a billion downloads a month.

What "stateless" means here

Old MCP held a conversation. The client and server shook hands, the server issued a session ID, and both sides kept that thread alive across every request that followed. That works fine on a laptop. It is miserable in production, because a session pinned to one server instance means you need sticky routing and a shared session store before you can safely run a second copy.

The new spec deletes all of it. No handshake, no session header, no thread to keep alive. Every request now carries what it needs, so any request can land on any instance behind an ordinary load balancer. @rohanpaul_ai described the shift as turning "a session-heavy connection into a simpler request-based protocol," and that is most of the story. MCP stopped being special and started behaving like the rest of the web.

The practical result is that a connector can run on serverless infrastructure and cost close to nothing when nobody is using it. Requests also carry routing headers now, so a gateway can decide who is allowed to call which tool without cracking open the message body, and directory listings can be cached with an expiry. None of that is exciting. All of it is what software needs before a vendor will run it for ten thousand customers.

The boring half is the important half

Six of the changes in this release are about authorization. Dynamic client registration gives way to signed metadata documents, credentials get bound to the server that issued them, and issuers get validated properly. There is also a formal extensions framework, so new capabilities arrive as versioned add-ons instead of bloating the core. Two shipped with it: MCP Apps, which lets a connector draw a real interface inside the chat window, and Tasks, for work that takes longer than a single request.

Put together, that is a protocol being fitted for companies that have a security team. @paulakshat flagged the direction ahead of the final release, writing a developer breakdown of why "MCP is going stateless" and what it costs the people maintaining servers.

My read

Something real broke here. If you built an MCP server that leaned on sessions, this is a rewrite, not a version bump. The spec ships with a deprecation policy promising a twelve-month minimum transition window, which helps, but it does not make the work go away. The reaction I have watched splits along a clean line: the people who operate this infrastructure are relieved, and the people who have to migrate their code are not.

The ecosystem did not wait around. GitHub's MCP server announced support days before the spec was even final, and both AWS and Microsoft published guidance the same week. That is what a standard looks like once it stops being one company's idea.

What this changes for a business around here

Nothing this week. Over the next year, quite a bit.

The reason "connect AI to our actual systems" has been a custom project is that the connector layer was expensive to host and awkward to secure. Both of those just got fixed. Expect your scheduling software, your invoicing tool, and your CRM to start shipping official connectors, the way they all eventually shipped mobile apps.

Which puts the burden somewhere unglamorous: your side. A connector can only reach data that lives in a system, so anything still sitting in a notebook, a shared inbox, or one person's head stays invisible to it. And permissions stop being a formality once software can act on its own. If everyone at your shop logs in as the same user, you will not like what that looks like when an agent is holding the keys.

That is the kind of thing we sort out in a free process audit at New Face Design: where your information actually lives, who is allowed to touch it, and which repeated work is worth handing to software before the connectors show up asking.