A Quiet Release With Big Consequences

If you read our earlier piece on custom MCP servers, you know the Model Context Protocol (MCP) is the standard “key card” that lets an AI assistant safely look at — and act on — the systems your business runs on, like your time & attendance, payroll, or CRM.

On July 28, 2026, the people who maintain MCP published the biggest revision to the standard since it launched in late 2024. The headline change is a single word: stateless. It sounds technical, and it is. But the effect is something every business owner can appreciate — AI connections that are more reliable, scale more easily, and cost less to run.

Here’s what changed and why it matters, without the jargon.

What “Stateless” Actually Means

Until now, every conversation between an AI assistant and an MCP server started with a handshake. The two sides introduced themselves, agreed on the rules, and the server then had to remember that conversation — a “session” — for as long as it lasted.

That memory was the problem. Picture a busy deli with three people behind the counter, where only the person who took your order can ever finish it. If they step away, you start over. If the line gets long, you can’t just send customers to whoever’s free.

That’s how MCP servers worked. When a company ran several copies of a server to handle more traffic, every request from a given user had to go back to the same copy, because only that copy remembered the session. If that copy restarted, the session was gone and users saw errors.

The new version removes the handshake and the session entirely. Every request now carries everything the server needs to know — which version of the protocol it speaks and what the client can do. Any copy of the server can answer any request. In deli terms: you hand over a complete order slip, and whoever is free fills it.

Stateless means no request depends on a server remembering the last one. Any server can pick up any request — so a restart or a busy moment no longer breaks the conversation.

Why That Matters to You

You’ll never see this change directly. You’ll feel it.

  • Fewer dropped connections. When a server restarts or updates, the next request simply goes to another copy. No lost sessions, no “please try again.”
  • It scales like an ordinary website. MCP servers can now sit behind the same standard load balancers that serve web pages. Google, writing about the update, described the old way as a common source of “session not found” errors when traffic was spread across servers.
  • Lower running costs. Servers no longer need extra infrastructure just to remember sessions. GitHub, for example, has dropped the session storage its MCP server used to depend on. Servers can also run on “serverless” platforms that shut down to zero when nobody is using them — so you’re not paying for idle machines overnight.
  • Faster responses. Lists of available tools can now be cached, so an assistant doesn’t have to re-ask a server what it can do on every turn.

For a small business, that adds up to AI that’s always on during the early shift and the weekend close, without the cost of keeping heavyweight servers running around the clock.

“Are You Sure?” Is Now Built In

One of the most practical changes is how a server asks for confirmation in the middle of a task.

Say a manager asks an assistant to post a schedule change, or approve a batch of missed punches. A well-built assistant should stop and confirm before it changes anything. Previously, that kind of back-and-forth required holding a connection open while waiting for the person to answer.

The new standard handles it more cleanly. The server replies with “I need more information,” lists exactly what it needs — a confirmation, a choice, a missing detail — and the assistant asks the person. Once they answer, the original request is simply sent again with the answer attached. Supabase, one of the early adopters, said this lets its tools confirm cost before creating a project and warn before running a query that would delete data.

That’s exactly the behavior you want from AI that touches payroll and timecards: it answers freely, but any change pauses for your approval first.

Tighter Sign-In Security

The update also hardens how AI tools sign in to the systems they connect to. In plain terms:

  • Clients now verify who issued their login. This closes a known loophole where a sign-in from one service could be confused with another.
  • Credentials are tied to the system that issued them and can’t be reused somewhere else.
  • Desktop and command-line tools register correctly, which fixes a common source of failed sign-ins.

The maintainers noted that sign-in is where builders spend most of their integration time, so tightening it at the standard level means fewer homegrown — and potentially weaker — workarounds.

Easier for Security Teams to Govern

Every request now carries simple labels on the outside saying what it is and which tool it’s calling. That lets firewalls, gateways, and rate limiters route, meter, and block traffic without opening up the request itself. If your IT provider or security team wants to allow some AI tools and not others, or cap how often a tool can run, that’s now straightforward.

What Goes Away — and How Long You Have

A few older features are being retired, including the ability for a server to ask the AI model to write text on its behalf, and a couple of logging and folder-sharing features. None of them disappear overnight: the project adopted a formal policy that deprecated features keep working for at least 12 months before removal, and there’s a public list of what’s on the way out.

The official software kits in the four most widely used programming languages — TypeScript, Python, Go, and C# — shipped support alongside the release, and major platforms including Cloudflare, Amazon Bedrock AgentCore, Microsoft Foundry, and Netlify announced support.

Do You Need to Do Anything?

If you’re using an AI assistant from a vendor, the work is on their side, not yours. The questions worth asking are simple:

  1. Are your MCP connections on the 2026-07-28 version, or on a plan to get there?
  2. Do actions that change data pause for confirmation?
  3. Are older, deprecated features being phased out before the 12-month window closes?

At CTR/NY, we design, build, and run the custom MCP servers behind our AI Workforce Assistant and Workforce Analytics. Keeping those connections current with the standard — so they stay reliable, scoped to each user, and confirmed before they act — is part of the service, not an upgrade project you have to manage.

The Bottom Line

The July 2026 MCP update doesn’t change what an AI assistant can do for your business. It changes how dependably and affordably it can do it. Stateless connections mean fewer errors and lower costs, built-in confirmations mean safer actions, and tighter sign-in rules mean fewer security gaps. It’s the kind of plumbing upgrade that makes AI feel less like an experiment and more like a utility.

Sources: the official MCP 2026-07-28 release announcement, the specification changelog, and Google’s Scaling AI Agent Infrastructure with the MCP Stateless Updates.