Your connectors just became infrastructure

Say you run revenue operations and you want ChatGPT to pull last week's closed deals out of your CRM, draft a follow-up for each one, and drop the drafts in your Gmail. For two years the honest answer has been the same: you can describe that in thirty seconds and you still can't build it, because getting the AI to reach into those three tools and act on what it finds requires an engineer to write the connective code and then keep that code from breaking every time a vendor changes something. Today the standard that does the connecting, Model Context Protocol, shipped its largest revision since it launched. It makes those connections cheaper to run, lets them show you a real interface instead of a wall of text, and, for the first time, promises not to break the ones you rely on without a year of warning. What that displaces is the integration ticket, and the contractor's invoice that used to come attached to it.

What actually changed

MCP is the plug standard for AI. It is the reason a chat assistant can list your calendar or search your Drive instead of just talking about them. If you have ever added a connector to Claude or wired a plugin into ChatGPT, you have used MCP without being shown the label. The finalized version, published today and dated 2026-07-28, is a real rewrite rather than a polish, and the specification blog lays out the whole thing in more detail than most people will ever need.

Four pieces of it matter to anyone building or buying these tools. The first is that the protocol went stateless. In plain terms, each request an AI makes to one of your connected tools now carries everything it needs inside it, so the connector no longer has to remember who it was talking to between calls. That sounds like an engineering footnote. It is actually the difference between a connector you can run cheaply on ordinary web hosting and one that needed special infrastructure to keep a conversation pinned to one server. The floor under running your own tools just dropped.

The second is a new extension called MCP Apps, which lets a connected tool render an actual interface, buttons, a form, a small interactive panel, right inside the assistant, in a sandboxed frame the host has to review before it runs. Until now a tool could only hand the AI text back. Now the thing you build can show a real surface. The third, called Tasks, lets a tool accept a long job, hand back a ticket, and run in the background while you go do something else, instead of forcing you to sit and watch a chat window for six minutes. The fourth is the least glamorous and possibly the most important: a formal deprecation policy, which means any capability the standard retires has to sit marked for at least twelve months before it can actually be removed, plus authorization that lines up with the ordinary OAuth login flow every business tool already uses.

That is the news, and at almost any other outlet that would be the whole article. Here it is the setup.

Why this matters if you are not an engineer

Start with who this is for, because it is not everyone. If your entire relationship with AI is opening ChatGPT in a browser and using the connectors that came in the box, today changes nothing about your Tuesday, and you can stop reading with a clear conscience. This is for the person one step past that: the operations lead who has started stitching the AI into the actual tools the business runs on, or is about to, or is paying someone to.

For that person, the quiet story of the last year has been that the wiring was the expensive part. The model got cheap and good. Describing what you wanted got easy. But the moment you wanted the assistant to do the thing, reach into HubSpot, touch QuickBooks, post to Slack, act and not just advise, you hit a wall that had an engineer on the other side of it. Every custom connection was a small software project. It had to be built, it had to be hosted somewhere, it had to be secured, and it had to be maintained, which is the word that quietly eats budgets, because maintenance means someone is on the hook the next time a vendor ships an update and the glue snaps. That is the ticket you file and wait six weeks on. That is the invoice.

Three of today's four changes go straight at that wall. Stateless connectors are cheaper and simpler to stand up and run, which lowers the bar for the tools you build yourself and for the small vendors selling you ready-made connectors, because their hosting bill just got smaller too. The deprecation policy is the one that should make an owner exhale, because it attacks the maintenance tax directly. The reason homemade integrations rot is that the ground moves under them without notice. A standard that guarantees a year of warning before anything it retires disappears is a standard you can build a small business process on without lying awake about it. And the authorization cleanup matters the day you point one of these tools at something that holds real data, your customer records, your inbox, your books, because it means the connection logs in through the same front door your other software uses, with the same permissions and the same audit trail, rather than through some bespoke arrangement nobody wants to explain to an auditor.

MCP Apps is the one that changes what a homemade tool can even be. A tool that can only return text is a tool you have to read and then go act on somewhere else. A tool that can show you a real panel, three deals with an approve button next to each, a form that fills itself from the record it just pulled, is a tool you can actually run your day through. That is the gap between an assistant that drafts and a system that does, and it is the gap most small businesses fall into when they try to turn an AI experiment into an AI process. HubSpot, to pick one example our readers actually use, already shipped a command-line connector built for exactly this kind of agent-driven work. Today's standard is the common shape all of those connectors are converging on, which means the ones you adopt will increasingly speak the same language instead of each being its own snowflake.

The honest take

Here is what the specification page will not tell you in the same font as the good news. You do not get any of this today, even though the spec published today. A standard is a promise about how software should behave, not the software itself. The actual connectors you use, the Claude connector, the ChatGPT plugin, the tool your vendor sells you, all have to be rebuilt to match, and the spec itself gives the toolmakers a ten-week window just to validate the changes against real workloads. The people doing that rebuilding are mostly running their own AI agents through it now rather than hand-writing every line, so it will move faster than these migrations used to. It will not be instant. If you are waiting to feel this, you are waiting on the tools you use to adopt it, and some will be quick and some will drag.

Second, this release breaks things on purpose. Anyone who built against the earlier, experimental version of Tasks has to migrate to a redesigned version, and there is a small change to an error code that will quietly break connectors that were checking for the old one. For you that work is invisible, which is exactly the problem, because invisible-to-you does not mean nobody has to do it. It means your vendor has to do it, and the honest question to ask any tool you are paying for is whether they are on top of the 2026-07-28 changes or whether your integration is going to have a bad month while they catch up. That is a real support conversation to have before the summer is out.

Third, MCP Apps is a genuinely new surface, and new surfaces are new places for trust to leak. A connector that can draw an interface inside your assistant is a connector whose author decides what that interface shows and what its buttons do. The standard routes every one of those button presses back through the same consent and audit path as a normal tool call, which is the right design, and the host is supposed to security-review the interface before it renders. But supposed-to is carrying weight in that sentence. The first time you approve a slick panel from a tool you half-remember installing, you are trusting whoever built it a little more than you trust a plain text reply. That is not a reason to avoid it. It is a reason to keep the same short list of tools you would let touch your money and your customers, and to not let a nice interface talk you into expanding it.

And to be square about scale: none of this is a frontier model launch, and it will get a fraction of the coverage the day's louder stories got. It is plumbing. But plumbing is the thing that decides whether the expensive water actually reaches the tap, and for the person trying to build their own tools without hiring an engineer to do it, the plumbing was the whole problem. Today it got a serious upgrade and, for the first time, a warranty.

Connectors spent two years as the weekend hack you were slightly embarrassed to depend on. As of today they are infrastructure, and infrastructure is the thing you stop thinking about right up until the morning it is quietly holding the whole business together.