MCP Just Had Its Biggest Update Ever. Here's What Changes For You

Your AI assistant has been quietly depending on plumbing you've never seen, and today that plumbing got rebuilt.

Priya runs a six-person physiotherapy clinic in Manchester. Three months ago she connected Claude to her booking calendar and her patient management system, so it could answer "who's free Thursday afternoon" without her opening three different apps. Most days it worked. Then, every few weeks, it just stopped. No error. No warning. Claude would say it couldn't find the calendar, and Priya would lose twenty minutes reconnecting something she never touched.

She assumed it was Claude. It wasn't Claude. It was the plumbing underneath.

What is MCP, and why does your AI depend on it?

Every time an AI agent reads your calendar, updates a customer record, or pulls last month's sales numbers, it's talking through something called the Model Context Protocol, or MCP. Anthropic introduced MCP in late 2024 as an open standard, and it has since become the plumbing that nearly every AI tool uses to reach outside its own head.

Think of it like this. Before MCP, connecting an AI model to your booking calendar meant a developer had to hand-write custom code: fetch the calendar's API documentation, wire up authentication, handle every edge case, and hope nothing on the calendar's end changed. MCP replaced that one-off wiring with a shared plug. Any AI model that speaks MCP can talk to any tool that offers an MCP server, the same way any laptop with a USB-C port can talk to any USB-C charger.

An MCP server sits between your AI and your actual tool. It translates "find Thursday's open slots" into whatever request your calendar software understands, handles the login with a scoped access token instead of your full password, and hands the answer back. You never see any of this. You just ask a question and get an answer, most of the time.

What changed on July 28, 2026?

Until today, every MCP connection started with a handshake: the AI and the tool's server introduced themselves, agreed on a session ID, and that session lived in one server's memory for as long as the connection lasted. If that specific server hiccuped, or the company running the tool quietly moved traffic to a different machine, the session vanished. Your AI assistant lost the thread. That is very likely what kept happening to Priya.

The new specification, dated 2026-07-28, removes the handshake and the session ID entirely. Every request now carries everything a server needs to answer it on its own, so any available server in a pool can pick it up. Nobody has to stay stuck to one machine. In plain terms, the connection between your AI and your tools stops depending on one fragile thread staying open, and starts working the way ordinary, boring, reliable web traffic works.

The second change matters just as much, and it has nothing to do with servers. Before this release, there was no formal rule about how much warning you would get before a feature in MCP was removed. A tool you depended on could change with no notice at all. As of today, any feature has to keep working for a minimum of twelve months after it is marked for retirement, and it has to be listed in a public registry first. Three developer-facing features, Roots, Sampling, and Logging, are the first to get that twelve-month clock. Nothing a business owner uses day to day stops working tomorrow.

Isn't this just plumbing that developers care about?

It's fair to ask. You didn't sign up to understand session handshakes. But you did sign up to trust an AI tool with your calendar, your customer records, or your stock count, and that trust has a failure mode you have probably already met: the tool randomly stops working, for no reason you can find, and you quietly stop trusting it.

That failure mode is exactly what today's update targets. You do not need to understand a request envelope to feel the difference. You will feel it as fewer mystery disconnects, and when something in your AI's toolbox does eventually change, you get a year of public notice instead of a surprise Tuesday morning.

There is a second, more exciting effect. Because MCP now works cleanly behind standard, ordinary infrastructure, more software makers are racing to ship their own MCP connectors instead of making you hire a developer to build one. InvoiceCave, a small invoicing tool, announced on X this week that it shipped 102 first-party MCP tools, so an AI agent can create invoices and track payments through plain language. DC Hub, an infrastructure data platform, announced 74 of its own the same week. Treat those specific counts as self-reported for now, but the pattern behind them is real and worth watching: the software you already pay for is increasingly likely to just appear in your AI assistant's connector list, with no custom integration required.

How do you actually use this today?

1. Check what you already pay for. Open your AI assistant's connector or "apps" settings, Claude and ChatGPT both have one, and search for the tools you already use: your calendar, your CRM, your accounting software, your booking system. If it has an MCP connector, it shows up there.
2. Connect one tool, not five. Start with whichever tool costs you the most manual re-typing today. For Priya that was the booking calendar. For a retailer it might be the inventory system. For a support-heavy business it might be the helpdesk.
3. Give it a narrow, low-risk first job. Ask it to summarize before you ask it to edit. "Tell me who's free Thursday" comes before "reschedule Thursday's appointments." Read access before write access, always.
4. Watch it for a week before trusting it unsupervised. MCP connectors carry scoped access tokens, meaning the AI only gets the access you explicitly grant it, but that protection only works if you actually check what you granted.
5. Expect fewer surprises from here. If a connector you rely on changes in the next year, the new deprecation policy means you will see it coming, not discover it the day the tool stops answering.

What's actually different, side by side?

What you have after this

A week from now, nothing about your day changes unless you go looking for it. A month from now, if you have connected even one tool, you will notice the small thing: the AI did not need reconnecting this time. Six months from now, once two or three of your everyday tools show up as ready-made connectors instead of custom projects, you will have an assistant that quietly holds your calendar, your customer list, and your numbers in the same conversation, without you writing a line of code or hiring anyone who does.

Same tools. Same tasks. A machine that finally stays plugged in.