Guides · Claude
The standard that connects AI assistants like Claude, ChatGPT or Copilot to your company's tools
Ironhack for Business
- Copy and paste between tools
- AI works with your data
MCP is the piece that lets AI work with your company's data instead of staying inside the chat window. In Spain, 21.1% of companies with 10 or more employees use artificial intelligence in the first quarter of 2025 (INE, survey on ICT and e-commerce use in companies 2025). Without a connection to company systems, that use depends on copying and pasting information between tools, and MCP is what changes that.
What is MCP?
MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems: databases, tools and workflows. Its official documentation compares it to a USB-C port: just as USB-C standardises how electronic devices connect, MCP standardises how AI connects to other systems.
It is a standard introduced by Anthropic in November 2024. Its practical advantage is that it cuts the time and complexity of integrating a tool with an AI assistant: the connection is built once and works with any compatible assistant.
These are examples from the documentation itself:
- An assistant that reaches your Google Calendar and Notion to work with your schedule and your notes.
- A corporate chatbot connected to several of the organisation's databases, to analyse data from the chat.
- Claude Code generating a complete web application from a Figma design.
How does MCP work?
MCP works with 3 pieces: the host, which is the AI application your team uses; the MCP client, which lives inside the host and translates requests; and the MCP server, which gives access to an external tool's data or actions. That is how the MCP architecture documented by Google Cloud describes it, with JSON-RPC 2.0 messages that can travel locally or remotely.
MCP · Architecture
The same connection serves any compatible assistant.
The application
Host
The AI application your team uses: Claude, ChatGPT or Copilot.
Inside the host
MCP client
Translates the assistant's requests into the protocol.
One per tool
MCP server
Gives access to an external tool's data or actions.
In practice the flow is this: someone on your team asks Claude «summarise the campaigns whose cost went up this week». Claude, acting as the host, uses its MCP client to call the ad tool's MCP server. The server returns the data and Claude answers with it, without anyone having to export a report.
What is an MCP server?
An MCP server is the service that gives AI access to a specific tool's data and actions — your CRM, your BI tool or your CMS. Many vendors publish their own. DatoCMS, for example, offers an official MCP server that can create records, upload images, add translations or inspect how content models are defined.
Each server defines what the AI can do, and your company decides what permissions it grants. DatoCMS's offers 3 levels: read content only, read and edit, or everything the user can do.
How is MCP different from an API or from RAG?
MCP lets AI query data and run actions in any tool that speaks the same standard, while an API connects 2 specific systems and RAG only retrieves information to improve answers. The difference with RAG is action: per Google Cloud's comparison with RAG, RAG is passive retrieval of information, while MCP is a two-way interaction that can, for example, update a CRM.
| MCP | API | RAG | |
|---|---|---|---|
| What it does | Queries data and runs actions in external tools | Connects 2 specific systems | Retrieves information to improve the answer |
| How it is built | A common standard: one server serves any compatible assistant | One integration per pair of systems | A document base the AI consults |
| Example | An agent that updates a CRM record | A web form that sends leads to the CRM | A chatbot that answers from your internal documentation |
Why has MCP become the standard?
MCP has become the standard because the main AI assistants have adopted it and it no longer depends on a single company. It has over 97 million monthly SDK downloads, 10,000 active servers and support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code (MCP Blog, MCP joins the Agentic AI Foundation, Dec 2025).
Since December 2025, MCP is part of the Agentic AI Foundation, a Linux Foundation body co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg.
For your company this means an MCP connection does not tie you to one assistant. Claude, Copilot or ChatGPT: the same MCP server works with whichever your team already uses.
How we use MCP at Ironhack for Business
We use MCP in 2 internal marketing processes. Neither started from the technology, but from a repetitive task we wanted to take off the team.
Marketing: Meta Ads alerts in Slack
At Ironhack, our marketing team has Claude connected to Meta Ads through MCP, with read-only permissions. Claude reviews campaign performance and sends alerts to a Slack channel when it spots an anomaly, for example:
- An ad that has not had enough budget to be tested.
- A cost per application that doubles.
- A CTR that halves.
The point is for the team to spend less time reviewing campaigns one by one and more deciding what to pause, what to keep and what to scale.
DatoCMS: updating the site from Claude
We update our website's content from Claude through the DatoCMS MCP server. What makes it work is the mapping done first: we keep files that say which CMS field each block of content belongs to. That way Claude knows where the title goes, or the meta description, or each section, without anyone explaining it every time.
In both cases, connecting the tool was the easy part. What makes the difference is the context: what data Claude needs, with what permissions, and what rules it follows.
What are the risks of MCP for a company?
The main risk of MCP is giving AI more access than it needs, because a broad permission turns any mistake or attack into a bigger problem. The MCP specification itself warns that overly broad permissions increase the impact of a stolen credential and make auditing harder.
These are the risks your leadership team should know about:
Excessive permissions
The specification recommends starting with minimum read permissions and widening them only when a task requires it.
Servers of unknown origin
A local MCP server runs with the same privileges as the AI application, so a malicious one can run commands or extract data from the machine.
Authorisation failures
The specification documents attacks such as the «confused deputy» or «token passthrough», where a badly implemented server lets a third party act on a user's behalf.
Shadow AI
People connecting MCP servers on their own, without approval or a record.
The basic measures are well known: a list of approved servers, read-only permissions by default — as in our Meta Ads connection — and human confirmation for actions that write or delete. DatoCMS, for example, asks the user to confirm before every use of its destructive tools.
If you do not know which AI tools circulate in your company today, our AI consulting audits which ones are in use, unapproved ones included, with what data and at what risk, and delivers data and AI governance with a usage policy, approved tools and a register. To review what documentary evidence you are missing against the AI Act, take our self-assessment test on Article 4.
How do you start using MCP in your company?
Start with one concrete process, not by connecting every tool at once. An MCP rollout that makes sense combines 3 layers:
MCP · Where to start
One process, its governance, and the team that will use it.
What to connect
Pick 1 or 2 repetitive processes with an owner and a KPI. Check whether the tools involved already publish an MCP server.
With what governance
Define which servers are approved, what permissions each has and which actions need human confirmation. Start read-only.
With what team
Technical people set up and review the servers; non-technical people learn to brief the AI well and review what it does.
Without the team layer, the connection exists but nobody uses it with judgement. For that part, our AI workshops include technical sessions for engineering teams (generating, refactoring, testing and pair programming), with Claude as the main tool and Copilot or ChatGPT depending on your company's stack. If you want a sustained programme, our in-house AI school includes a 1:1 leadership level where each director works with a personal agent connected to their data and their email.
Is MCP free?
Yes, MCP is an open-source standard, so using it carries no licence cost. What you do pay for are the tools you connect and the licences for the AI assistant your team uses.
Does Claude MCP also work with ChatGPT or Copilot?
Yes, an MCP server works with any compatible assistant, and ChatGPT, Claude, Gemini and Microsoft Copilot are compatible (MCP Blog, MCP joins the Agentic AI Foundation, Dec 2025). That is why an integration you build for Claude is not lost if your company switches or combines assistants.
Do I need to know how to code to use an MCP server?
Not to use a server that is already published: many are authorised with a sign-in, like the DatoCMS one, which uses OAuth authentication. You do need a technical profile to build your own server, or to review the security of one before approving it.
Next step
Pick 1 process where your team copies data from one tool to another every week, and check whether that tool already publishes an MCP server. If you want to do it with a method and the right governance, we will come back with a starting point in under 1 working day.