What is an MCP, and what can it do for your operation?

MCP stands for Model Context Protocol. You've probably come across it in a headline, in the release notes of a tool you already use, or in any conversation about artificial intelligence, almost always without an explanation attached.

Adoption is moving fast, including in the tools a delivery operation runs on: Google Maps, online stores like Shopify and WooCommerce, and Business Central, Microsoft’s ERP, all have their own MCP.

This article explains what it is, what it's for, and what it doesn't do. No jargon.

What is an MCP?

An MCP is the standard way for an AI assistant, such as Claude, ChatGPT or Gemini, to use a piece of software you already work with.

It isn't a robot. It isn't a model. It's the cable between the two.

The acronym stands for Model Context Protocol, and the name is the least interesting part. What matters is what changes: until recently, an assistant only knew what you pasted into the chat. With an MCP it goes into the system and looks up the real data.

For you that turns into something concrete: you stop going to find information and start asking for it.

Why is adoption moving so fast?

Anthropic published the protocol in November 2024 and opened it to anyone. OpenAI adopted it the following year. In December 2025 it stopped belonging to a single company and moved to the Linux Foundation, with Google, Microsoft and AWS among its backers.

The numbers back it up. In September 2026, Anthropic reported that MCP usage across its products was up 110x so far that year.

July 2026 also brought a new version of the standard, the one many people call MCP 2.0: it simplifies how an MCP is built so it handles more load and is easier to maintain. It has stopped being an experiment and is being prepared to last.

In other words: this isn't one supplier's bet. It's the plug everyone agreed on.

If you work in transport, you know this story. Until the 1950s, ships were loaded by hand, one bundle at a time, and every port moved cargo its own way. In April 1956, Malcom McLean loaded 58 containers onto a converted tanker and sent it from New Jersey to Houston. Loading took under eight hours, where it used to take days. The container didn’t make ships any faster. What changed is that everyone started using the same box, with the same dimensions: any crane could lift it and any truck could carry it, with nothing to adapt.

An MCP is that, applied to software. It’s the standard box: if a program has an MCP, any assistant can use it without a custom-built integration.

What can you ask it?

It's easiest to see in real sentences. In a delivery operation they sound like this:

• "Which stops failed yesterday, and why?"
• "Import this order spreadsheet into tomorrow's plan"
• "Import these 340 stops into tomorrow's plan"
• "How many routes missed their time window last week?"
• "Re-optimize tomorrow's plan without touching what's already assigned"
• "Summarize how yesterday went before I get to the office"

There's no new screen to learn. You type into the assistant the way you'd message a colleague.

Asking isn't the same as doing

Look at that list again. The first line asks a question. The second one changes things.

MCPs do both, and it pays to be clear about it from the start: reading looks things up and touches nothing; writing creates, edits or launches a process. You can start with reading alone.

What an MCP doesn't do

It does not put your operation on autopilot, whatever you may have read.

It isn't the intelligence: the model is a separate thing. It doesn't replace your systems, and each one still covers its own half of the journey: the data stays where it was, under the same permissions. And it won't fix bad data. If your history is messy, the wrong answer just reaches you sooner.

Routal's MCP

Routal has one, inside its analytics and integrations layer. You connect your account to an assistant like Claude and from there you can ask about your projects, review how yesterday's deliveries went, search stops, import tomorrow's, or launch an optimization.

It works on your real operation, with your user's permissions. Whatever you can't do in the platform, you can't do here either.

Our favourite part isn't the AI. In almost every operation there's one person who knows where to look: where the history sits, which report answers which question. None of it is written down. When asking is as easy as typing a sentence, that knowledge stops living in one head.

How to connect it

Two minutes, and nothing to install:

1. In your assistant, open the connector directory and search for Routal.
2. Hit connect and sign in with your Routal account.
3. Approve the access and you're done.

You can disconnect it from the same place whenever you want.

How to tell a trustworthy MCP

This goes for Routal's and for anyone else's. Three things to check before connecting one:

• It comes in as you. It inherits your permissions and not one more. If it asks for a master key, think again.
• You know what it can touch. Reading and writing aren't the same thing, and you decide where to start.
• You can cut it off. Revoking access should be a button, not an email to support.

They're the same conditions you'd set for any new user.

Try it with a single question

Nothing about the way you work today has to change. You connect it, ask your first question and see what comes back. If the answer holds up, that's one screen fewer to open every morning.

From there, you use it for whatever earns its place.

Want to see it with your own data? Get in touch and we'll try it with your operation on screen.