The new front door: what changed in MCP last week

When your next customer arrives through an AI assistant instead of your website, will your business be there to meet them?

For twenty years the playbook for reaching customers online has been stable. Build a website, optimize it for Google, maybe add an app. The interface belonged to you, and search engines delivered the traffic.

That is changing. People increasingly start inside an assistant: find me winter tires for a Golf, compare these two insurance offers, book a service appointment next week. The assistant researches, narrows the options, and often completes the task itself. No browser tab involved.

Which brings us to the Model Context Protocol, and to why last week's revision matters more than protocol housekeeping usually does.

What MCP actually does

MCP is an open standard for connecting AI assistants to external systems: product catalogs, booking systems, customer databases, internal tools. A connected assistant can check your stock, hold a slot, place an order. An unconnected one can only describe you from the outside. Anthropic introduced the protocol in late 2024, and it has since become the de facto standard. Claude's connector directory alone lists over 950 servers, with millions of daily users.

Ask an assistant for a family car under 400,000 kroner, available nearby, and the quality of the answer depends entirely on whose inventory it can reach. Dealers with a connector get shortlisted. That is the whole mechanism.

What changed in the new spec

The 2026-07-28 revision is the fifth release of the protocol. Three things in it remove the objections we run into most often when connecting a company's systems to AI.

The big one is that MCP went stateless. Servers no longer hold long-running sessions, so they behave like ordinary web services and run on serverless and edge infrastructure without special handling. That reads like a footnote until you price it out: connecting your systems to an assistant is now ordinary web development, with ordinary hosting and ordinary scaling, which is a very different conversation to have with whoever owns your infrastructure.

MCP Apps, the first official extension under the new versioned framework, adds real UI. A business can render interactive components inside the conversation, so a customer configuring a product or picking a delivery slot is no longer doing it through plain text. Your brand gets to appear inside the assistant, with its own layout and behaviour.

Authorization was the other blocker. It now aligns with OAuth 2.0 and OIDC, the same standards behind Entra and Okta, so organizations can provision AI connectors through the identity systems they already run, with real access control. That shortens a compliance conversation that used to stall these projects for months.

There is also better observability for published connectors, and a research preview of MCP Tunnels for reaching systems inside private networks. Neither feature is aimed at people experimenting. Both assume something already in production, with people depending on it.

Why this should change your priorities

The pattern is familiar. In the 2000s you needed a website; in the 2010s you needed to load fast on a phone and rank on Google. The current requirement is being reachable by AI assistants, and it is less forgiving than either of those.

A business that skipped mobile still got visited, just awkwardly. A business that is absent from the AI layer can still be found through web search, but it arrives in the conversation as a citation. When the decision is actually being made, the assistant recommends whoever it can transact with.

The good news is that the work got smaller. With a stateless protocol and standard authentication, a well-scoped MCP server on top of systems you already run is a matter of weeks.

What to do now

Where we would start:

1. Map your entry points. The journeys that begin with a search or a question are the ones that move into assistants first.

2. Work out what an assistant would need to answer well: product data, availability, prices, booking. Most companies already hold all of it, so the gap is exposure, not data.

3. Build one connector, for your highest-value journey, end to end. A week of real usage will teach you more than a quarter of strategy work.

4. Treat it as a channel, not a project. Someone should own how the business shows up in AI interfaces, the way someone owns the website today.

This is the work we do at Nuto: taking systems a company already runs and making them usable by AI, from first sketch to something live. If you want to think out loud about what an entry point would look like for your business, we are easy to reach.