What is MCP? The Model Context Protocol is an open standard, announced by Anthropic on 25 November 2024, that gives AI applications one fixed way to connect to your data and tools [S-001]. One MCP server per system replaces a separate integration per AI tool. It matters once an AI assistant must read or act on your own data.
The official documentation compares it to a USB-C port: one standard socket that any compatible AI application can plug into, instead of a custom cable per device [S-002]. This article explains how that works, what you can do with it, and when you do not need it.
How MCP works: servers and clients
MCP has two roles [S-001]:
- The MCP server sits on the side of your data or tool. It exposes what is there: files in a folder, tables in a database, a CRM, a calendar. At launch Anthropic published ready-made servers for Google Drive, Slack, GitHub and Postgres, among others [S-001].
- The MCP client lives inside the AI application. Claude, but also ChatGPT, Visual Studio Code and Cursor support MCP [S-002]; OpenAI’s API does so through the Responses API [S-003]. The client asks the server what it can do, then uses that during the conversation.
The consequence: build one MCP server for your CRM and you can use it with every AI application that speaks the protocol. Switching AI vendors does not mean rebuilding the connection. That is the problem Anthropic set out to solve: every new data source needed its own implementation, and that did not scale [S-001].
What can a business do with it?
Run an AI assistant that does not just talk, but looks and acts. Three examples any small business will recognise:
- Ask questions of your own data. “Which clients have not ordered this quarter?” The assistant reads your CRM or database through an MCP server and answers, instead of someone exporting and filtering.
- Use the documents you already have. An assistant that can open your proposals, process descriptions or product information answers from your material rather than from general knowledge.
- Take actions within limits. Create a task, schedule a meeting, draft an email. The documentation mentions calendars and Notion as examples of what an agent can reach through MCP [S-002].
The difference from a chatbot that only answers is exactly this: access to systems. What an AI agent does with that access, and where the line is, is covered in What is an AI agent and what can it do for a marketing team?
MCP or a regular integration? The decision table
This is where most explanations of MCP go wrong: the standard does not replace your existing automations. A workflow in n8n or Make and an MCP server solve different problems. The question is who takes the initiative.
| Question | Regular integration (n8n, Make, Zapier) | MCP server |
|---|---|---|
| What starts it | an event or a schedule: new form, every Monday | a person (or agent) asking something during a conversation |
| What happens | a fixed sequence of steps, the same every time | the AI decides per question which data or action it needs |
| Predictability | high: you know in advance what will happen | lower: the answer depends on the question |
| Typical use | routing leads, reports, invoices | ”look this up”, “create this”, “summarise this” |
| Maintenance | workflow in the tool, by the team | server plus permissions, by whoever manages the systems |
| Needed when | the work is the same every time | someone wants to work with your data ad hoc through an assistant |
Rule of thumb: if the work is identical every time, build a workflow. If someone wants to ask your systems questions or have things done while working, MCP is the layer that makes it possible. Often you want both: the workflow does the fixed work, the assistant with MCP handles the exceptions. Which tool to use for the fixed work is covered in n8n vs Zapier.
What do you need to get started?
Less than it seems, provided you start small.
- One system, one purpose. For example: the assistant may read client data from the CRM. Not write, not the books, not everything at once.
- An MCP server for that system. For many common tools these already exist; otherwise it is build work for a developer or an agency, not a months-long project.
- An AI application with MCP support. Claude and ChatGPT have it [S-002]; on the work floor that is usually the desktop app or an internal chat window.
- Permissions. A separate account for the assistant with only the access the purpose requires.
If you already use Claude for text, you know the step: Using Claude AI for copywriting describes what the model does well and badly. MCP adds your own data to that.
Which risks do you need to cover?
Two, and both are organisational rather than technical.
Access. An MCP server gives an AI application access to a system. That access should be as limited as a new employee’s: read-only where reading is enough, only the tables that are needed, and its own account that you can revoke. Anthropic names “secure, two-way connections” as the goal of the standard [S-001]; it only becomes secure through how you set the permissions.
Actions. An assistant that can write can also write the wrong thing. Start with reading. Allow write actions (creating a task, changing a record) only once you have seen how the assistant handles read questions, and then with a confirmation step. It is the same rule as for any automation: the system does the execution, people keep the decisions. More on that boundary in automation versus AI automation.
Frequently asked questions
Is MCP only for developers? Building a server is developer work. Using it is not: someone using an AI application with MCP support simply notices the assistant can now look into the company’s own systems.
Does MCP only work with Claude? No. The standard is open; the official documentation lists ChatGPT, Visual Studio Code and Cursor alongside Claude as supporting applications [S-002], and OpenAI’s API supports MCP through the Responses API [S-003].
Does MCP replace my n8n or Make workflows? No. Workflows do fixed, repeated work on a trigger; MCP gives an assistant access during a conversation. They live side by side, see the decision table above.
How do I start without opening everything up? One system, read-only, its own account. Expand only once that has worked well for a few weeks.
Sources
- Anthropic, Model Context Protocol announcement, 25 November 2024 [S-001]
- Model Context Protocol, official documentation, introduction [S-002]
- OpenAI, API documentation on MCP [S-003]
An assistant that takes steps on its own and writes back into your tools is an AI agent. Where data only has to move from A to B, workflow automation is enough.
Want to find out whether an assistant with access to your own data solves something in your business, or whether a workflow is enough? Book a call. Thirty minutes, no preparation needed.