Skip to content
SciTechBiz Institute
I want to join
Blog · Forward Deployed Engineer

MCP explained for people deploying AI in companies

MCP (Model Context Protocol) is an open protocol that connects an AI agent to a company's tools and data. Each system exposes its tools on an MCP server and the agent uses them without a separate integration for each one. For whoever deploys it, what counts is which permission each tool has and what risk it brings.

Updated: 5 November 2026 · Deibson Silva · Forward Deployed AI Engineer and Founder

What it is

What is MCP, and which problem does it solve?

A language model, by itself, writes text. To do a job in a company it needs tools: reading a document, looking up a client in the management software, creating a task, sending a notice.

Until recently, each of those connections was built to order, and each AI application did it its own way. Changing application or adding a system meant redoing the work.

MCP, short for Model Context Protocol, is an open protocol that sets a common way. On one side is the MCP server, which exposes a set of tools, each with its name, its description and the data it expects to receive. On the other is the application the agent works in, which connects to the server, sees which tools exist and calls them when it needs them.

The idea comes down to this: one protocol, many tools. What is built once serves more than one agent.

Comparison

What changes between a custom integration and an MCP server?

Both connect an agent to a system. What changes is how much is reused and where control sits.

Custom integrationMCP server
How it connectsIts own code for each agent and each systemA common way: the server describes its tools and the agent discovers them
What is reusedLittle: each project starts from scratchThe whole server, in another agent or at another client with the same system
Where the permissions areScattered through the codeOn the server, tool by tool
How it is testedTogether with the agent, all at onceThe server first, without the agent; then the two together
What it asks of the builderKnowing each system from the insideThe same, plus deciding carefully what gets exposed
Tools

Which kinds of tools are exposed?

The tool is the agent's interface. It helps to separate them by what they can do.

Data tools

They read: check an order, find a document, bring back a client's history. They change nothing in the system.

Action tools

They write, send, pay or delete. They are what gives the agent its value and what calls for most care, above all when what they do cannot be undone.

Orchestration tools

They call another agent or another process. Useful when the work is shared out, and only after checking that a single agent is not enough.

Each one, with its risk

Low, medium or high, according to what the tool can do and whether it is reversible. That classification is written on the tool sheet before anything is connected.

Step by step

How do you build an MCP server with judgement?

The order taught in the Forward Deployed Engineer programme:

  1. 1Start from the process. Write down what the agent must be able to do in that job, and nothing more.
  2. 2Define one tool per action, with a clear name, specific parameters and error messages the agent can understand.
  3. 3Give each tool least privilege. If it only needs to read, it only reads.
  4. 4Classify the risk of each one and note it.
  5. 5Put prior human approval on every high-risk action, or block it.
  6. 6Test the server before the agent. If a tool returns more data than it should, this is where it shows.
  7. 7Log every call. Without a log, nobody can later explain what the agent did or why.

In the programme's lab, the server exposes at least three tools with their schema and their risk, and the high-risk one requires approval.

Risks

Which risks come with connecting tools to an agent?

The best known is indirect prompt injection. The agent reads a document, an email or a tool's response, and inside is a sentence written to make it do something else: forward a file, change a piece of data, skip a rule.

An agent without tools that falls for it writes a wrong answer. One with action tools can do something. That is why least privilege and prior approval come before convenience.

The second risk is exposing too much. A server that offers twenty tools when the process calls for four opens doors nobody watches.

The third is taking someone else's server on trust. Before you connect a server you did not write, you review which tools it brings and with which permissions, just as you review any code you did not write.

When a tool can act, someone has to be able to stop it. Article 26 of the EU AI Act requires deployers to assign human oversight of a high-risk system to people with the necessary competence, training and authority.

When

Do you need MCP in every project?

No. The rule of the method is to start with the simplest thing that works. Many processes are solved with a fixed flow, step by step, with no agent and no protocol.

An agent makes sense when there are complex decisions, rules that are hard to maintain or unstructured data. And a single agent comes before several.

MCP pays off when the same tools will serve more than one agent, or more than one company with the same system. In the method, what can be reused is documented once and goes into a library.

The protocol is still changing. Before you build, you read the current version of the specification.

Good to know

Honest limits

  • This article explains the idea of the protocol. The technical detail is in its specification, which changes over time.
  • MCP puts order into how tools are connected. What each one can do is in the hands of the person who builds the server.
  • A well-built server does not make an agent safe by itself. Tests with real cases and a simulated attack are needed too.
  • The programme's labs use synthetic data only. In the real project, the company's data never leaves its systems.
  • Today the programme is taught in Spanish: 27 Spanish videos are published at the Academy, and the English videos arrive in November 2026.
  • The certification attests a competence, never a job or an income.
FAQ

What people ask most

01

What does MCP stand for?

Model Context Protocol. It is an open protocol for connecting AI applications to external tools and data in a common way, instead of building a separate integration for each application and each system.

02

Does MCP replace the integrations a company already has?

Not necessarily. An MCP server usually builds on them: it uses the connections the system already offers and presents them to the agent as tools with a name, a description and permissions.

03

Is it safe to connect an agent to company systems with MCP?

It depends on how it is built. Safety lies in least privilege for each tool, in classifying its risk, in human approval for high-risk actions and in attacking the agent on purpose before release.

04

Do you need to know how to code to build an MCP server?

You need to be able to read and review code. In the Forward Deployed Engineer programme you build with a coding agent as a copilot: you decide which tools exist and with which permissions, and you check that they do what they should.

05

Where do you learn to build with MCP?

In the agent architecture module of the Forward Deployed Engineer Certification, with a lab in which you build an MCP server, connect it to an agent and complete the tool sheet.

Portrait of Deibson Silva
Deibson Silva · Forward Deployed AI Engineer and Founder

Founder of Legathum and iPersonality and Forward Deployed AI Engineer. He leads the SciTechBiz Community from Madrid and examines the FDE Certification defence.

Your licence

Start from your dashboard

On the licence page you request access; requesting access costs nothing. You then sign the contract and pay the annual licence from your dashboard.

iPersonality Global Institute, S.L. · NIF B27580042 · Madrid · Privacy policy

Prices in euros, exclusive of VAT. In Spain, Spanish VAT is added. In the rest of the EU, VAT is calculated at payment according to your country and VAT number. No Spanish VAT is charged to UK customers.