← All stories

What is MCP? The Model Context Protocol, explained for business teams

MCP is the standard that lets AI assistants use your systems. What it is, why it matters now, and how to start with safe, read-only tools.

This story is also available in Español

What is MCP? The Model Context Protocol, explained for business teams

Key takeaways

  • MCP, the Model Context Protocol, is an open standard for connecting AI applications such as Claude and ChatGPT to data sources, tools and workflows.
  • Since December 9, 2025 MCP has had neutral governance under the Linux Foundation's Agentic AI Foundation, so no single vendor controls it.
  • Use a plain API for fixed steps and an MCP server when an assistant must choose among tools or when several assistants need the same tools.
  • MCP connects, it doesn't restrict. Start with read-only tools and require approval or caps on anything that changes records or spends money, as Walmart does.

MCP, the Model Context Protocol, is an open standard that lets AI assistants like Claude and ChatGPT connect to your data, tools and workflows. For your team, the question isn't whether to adopt it. It is which systems an agent may read, which it may change, and who approves.

What is MCP in plain terms?

The MCP documentation calls it an open-source standard for connecting AI applications to external systems: data sources such as files and databases, tools such as search engines and calculators, and workflows such as specialized prompts. Its analogy is a USB-C port for AI applications: one standard connector instead of a custom cable for every device.

In practice, a company or software vendor publishes an MCP server that exposes a few actions, for example "look up an order" or "check availability." Any AI application that supports MCP can then use them. The docs list Claude, ChatGPT, Visual Studio Code and Cursor among the clients that do.

A giant robot reads a short menu at a restaurant table while a waiter heads for a small hatch in a closed kitchen door.
The analogy: the AI assistant is a diner, the MCP server is the waiter and the menu, and your systems are the kitchen. The diner can only order what is on the menu.
How it really works: the assistant can only call the tools the server exposes. Anything that changes a record goes through an approval step.

Here is an illustration. A hotel could publish an MCP server with two actions: check room availability and look up a booking. A support agent in any compatible assistant can then answer "is there a room on Friday?" without anyone pasting data into a chat window.

Why does it matter now?

Because no single company controls it. On December 9, 2025, the Linux Foundation announced the Agentic AI Foundation with MCP as one of its founding projects. Anthropic described its donation as a directed fund co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg, and said MCP's governance model stays unchanged.

For a buyer, that is the point. A standard backed by competitors is less likely to disappear or lock you in than one owned by a single vendor.

A real example: Walmart's Agent Gateway

On October 21, 2026, Walmart announced Agent Gateway, an MCP server that lets approved AI agents search, build carts and check out instead of navigating the site in a browser. It starts as a beta for companies and developers, per The Next Web.

The design is worth copying. Each agent gets its own profile inside the customer's account. The customer can approve every purchase or set a spending limit, such as $50, $100 or $200. Agents never see card details and can't add or change payment methods, and agents that break the terms are blocked.

Walmart didn't hand agents a key to the store. It gave them a narrow, capped door that a customer controls.

Simplified from Walmart's description. Purchases need customer approval by default, and Walmart may ask for extra approval even within a limit.

Notice what Walmart's design does not depend on: trusting the agent. The limits, the approvals and the hidden payment data work even when the agent gets it wrong. Use that test for any system you expose.

MCP or a plain API: which do you need?

Our rule of thumb: if the steps are fixed, use a plain API call or a workflow step. If an AI assistant has to choose which tool to use, or you want the same tools to work from several assistants, put an MCP server in front.

You needUse
A fixed sequence of stepsAPI call or workflow step
An assistant that picks among several toolsMCP server
The same tools in more than one AI appMCP server
A rule of thumb from our own projects, not a standard.

Most automations, such as "when an invoice arrives, extract the fields and post them to the ERP," don't need MCP. An agent answering a customer who might ask about an order, a refund or a booking does.

You don't have to choose one or the other. If you already run workflows in a tool like n8n, a workflow can sit behind an MCP tool: the assistant decides when to call it, and the workflow handles the fixed steps.

How to start this month

Start small, with a handful of read-only tools, and widen the scope only after you have read the logs.

  1. List the lookups an agent could do safely: order or booking status, availability, policy and FAQ search.
  2. Give it read-only access first, with credentials limited to those lookups.
  3. For anything that changes a record or spends money, require approval or a cap, like Walmart's.
  4. Ask each software vendor whether they offer an MCP server and what it can do.
  5. Read the security best-practices page in the MCP docs before you connect anything that holds customer data.

Pitfalls to avoid

  • Connecting everything at once. Every tool you expose is another way for the agent to go wrong. Start with three.
  • Write access on day one. Reading a wrong order status is an annoyance. Issuing a wrong refund is a loss.
  • Skipping the log. If you can't see what the agent called and when, you can't fix it. Keep a record of every tool call.

Sources

Which of your systems should an agent be able to use?We map the safe first tools and the approvals around them.

Map my operation

Frequently asked questions

What does MCP stand for?
Model Context Protocol. It is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows.
Who runs MCP now?
Anthropic donated MCP to the Linux Foundation, and it became one of the founding projects of the Agentic AI Foundation, announced on December 9, 2025. Anthropic, Block and OpenAI co-founded it, and Anthropic said MCP's governance model stays unchanged.
Do I need MCP to automate my business processes?
Not for fixed workflows, where an API call or a workflow step is enough. MCP helps when an AI assistant must choose among several tools, or when you want the same tools to work from more than one assistant.
Is MCP safe?
MCP standardizes how an AI app connects, not what it is allowed to do. Safety comes from what you expose and how you scope it: start read-only, limit credentials, and require approval or caps for actions that change records or spend money.
KytoWe design and build AI automations and business operating systems. Agency results + Academy sovereignty.

More stories

A black and white office facade at dusk with one window still lit in mint

Find the work your team could hand off.

Book 30 minutes. We map the workflows worth automating first and estimate the impact before you commit.