MCP (Model Context Protocol): what it is and how to use it

MCP (Model Context Protocol) is an open standard for connecting artificial intelligence assistants to a company’s systems and data — the ERP, the CRM, the ticketing tool, the documents — in an organized and controlled way. Launched by Anthropic in late 2024 as an open standard, it has been adopted by several AI companies. Instead of every assistant needing its own integration with every system, MCP defines a common way for them to talk.

In this article, I explain the problem MCP solves, how it works in plain language, what you can already do with it and the precautions before adopting it.

The problem: AI can’t see your company

Assistants like ChatGPT and Claude are great at writing, summarizing and reasoning. But by default they know nothing about what happens inside your company: they don’t see today’s orders, the status of a ticket, a customer’s contract or a product’s stock.

Until recently, connecting AI to these systems required a specific integration for each combination of assistant and system. Switched assistants? Redo it. Added a new system? Another integration. It’s the kind of work that drives up costs and stalls adoption.

What MCP solves

MCP does for AI what the standard power outlet did for appliances: it defines a common plug. A system that “speaks MCP” can be used by different compatible assistants, and vice versa (each application supports part of the protocol’s features). MCP builds on the systems’ APIs, which I explain in Systems integration: how to make everything talk.

In practice, this means:

• Fewer integrations to maintain. You expose the system once and it serves several assistants.
• Freedom to switch AI. The investment in the integration isn’t locked into one vendor.
• A growing ecosystem. Many well-known tools already offer an MCP connection, and the list grows fast.

How it works, without jargon

There are three pieces:

1. The AI application (the host, in technical terms) — the assistant the person uses (on the computer, in the browser, inside another tool). Inside it, a connector (the MCP client) talks to each server.
2. The MCP server — a small program that sits “in front of” a company system and tells the assistant what it can do there.
3. The system — the ERP, the CRM, the database, the documents folder.

The MCP server offers the assistant three kinds of things:

• Tools: actions the AI can perform, such as “look up an order”, “open a ticket” or “update the status of a deal”.
• Resources: information the application gives the AI as context, such as a document, a list of customers or a project’s history.
• Prompts: ready-made scripts the user chooses, such as “generate the team’s weekly summary” or “prepare this customer’s proposal”.

When the person asks the assistant for something, it works out which tools and information it needs, asks the MCP server and uses the answer to do the work — always within what the server allows and, for sensitive actions, only after your confirmation.

Examples of what you can do

• Project management: “Which of the team’s tasks are overdue, and who owns them?” — the assistant checks the task management tool and replies with the list.
• Sales: “Summarize this customer’s history before the meeting” — the assistant reads the CRM, the logged emails and the proposals.
• Customer service: “What’s the status of order 4521?” — the assistant checks the order system, instead of someone opening three screens.
• Operations: “Generate this week’s sales report and flag anything unusual” — the assistant queries the system and analyzes the numbers.
• Software development: the assistant reads the task in the management tool, checks the internal documentation and helps implement it.

Precautions before adopting it

Connecting AI to a company’s systems is powerful precisely because it starts to act — and that’s why it calls for care.

• Minimum permissions. The MCP server should expose only what’s necessary. If the AI only needs to look up orders, it shouldn’t be able to delete any.
• A human in control. The MCP specification itself recommends that there always be a person able to deny a tool from running. Sensitive actions — sending, deleting, paying — should ask for confirmation.
• Each person’s access. The AI should see only what the user making the request could already see in the system.
• Trusted servers. An MCP server performs actions in your systems. Use those built by trusted vendors or by your own team, and treat the rest with the same caution as any third-party software.
• Watch out for hidden instructions. An email or document read by the AI may carry malicious instructions (so-called prompt injection). That’s why human confirmation and minimum permissions matter so much.
• Log everything. Every action the AI performs needs to be logged: who asked, what was done and when.
• Data and privacy. What the server hands to the AI may be sent to the model provider. The same care with personal data applies here — I explain it in Artificial intelligence in business: where to start.

Where to start

1. Pick a question the team asks every day that today requires opening a system and searching. It’s the best first case.
2. Start with read-only access. Looking things up is much safer than changing them. Actions come later, with confirmation.
3. Check whether the system already has a ready-made MCP server. If it doesn’t, building a custom one is usually a lean project when the system already has a good API. The effort depends on permissions, access control and action logging.
4. Test with a small group and measure the time saved.

If your company is still taking its first steps with AI, the general roadmap is in the AI article mentioned above.

How WorkCodes uses MCP

We use MCP in our own day-to-day work: we built an MCP server for Jira that lets the assistant look up and organize the team’s tasks without leaving the conversation. It’s the same kind of bridge we can build for your company’s systems, following the precautions in this article: minimum permissions, confirmation for sensitive actions and action logging.

When it makes sense, we build the MCP server custom-made for your system. And if the question is still where AI pays off, our technology consulting helps you pick the first case. We work with companies remotely.

Get in touch and tell us which system you’d like to connect to AI. After an initial, no-strings conversation, you get a proposal with scope, timeline and price within 24 hours.