Introduction
in September 2025, I took half in a hackathon organized by Mistral in Paris. All of the groups needed to create an MCP server and combine it into Mistral.
Although my staff didn’t win something, it was a incredible private expertise! Moreover, I had by no means created an MCP server earlier than, so it allowed me to realize direct expertise with new applied sciences.
Because of this, we created Prédictif — an MCP server permitting to coach and take a look at machine studying fashions instantly within the chat and persist saved datasets, outcomes and fashions throughout completely different conversations.
On condition that I actually loved the occasion, I made a decision to take it a step additional and write this text to offer different engineers with a easy introduction to MCP and in addition provide a information on creating an MCP server from scratch.
If you’re curious, the hackathon’s options from all groups are right here.
MCP
AI brokers and MCP servers are comparatively new applied sciences which can be presently in excessive demand within the machine studying world.
MCP stands for “Mannequin Context Protocol” and was initially developed in 2024 by Anthropic after which open-sourced. The motivation for creating MCP was the truth that completely different LLM distributors (OpenAI, Google, Mistral, and so forth.) supplied completely different APIs for creating exterior instruments (connectors) for his or her LLMs.
Because of this, if a developer created a connector for OpenAI, then they must carry out one other integration in the event that they needed to plug it in for Mistral and so forth. This strategy didn’t enable the straightforward reuse of connectors. That’s the place MCP stepped in.
With MCP, builders can create a device and reuse it throughout a number of MCP-compatible LLMs. It leads to a a lot easier workflow for builders as they now not must carry out extra integrations. The identical is appropriate with many LLMs.
For data, MCP makes use of JSON-RPC protocol.
Instance
Step 1
We’re going to construct a quite simple MCP server that can have just one device, whose aim will likely be to greet the consumer. For that, we’re going to use FastMCP — a library that enables us to construct MCP servers in a Pythonic approach.
To begin with, we have to setup the setting:
uv init hello-mcp
cd hello-mcpAdd a fastmcp dependency (this may replace the pyproject.toml file):
uv add fastmcpCreate a most important.py file and put the next code there:
from mcp.server.fastmcp import FastMCP
from pydantic import Area
mcp = FastMCP(
identify="Whats up MCP Server",
host="0.0.0.0",
port=3000,
stateless_http=True,
debug=False,
)
@mcp.device(
title="Welcome a consumer",
description="Return a pleasant welcome message for the consumer.",
)
def welcome(
identify: str = Area(description="Identify of the consumer")
) -> str:
return f"Welcome {identify} from this wonderful software!"
if __name__ == "__main__":
mcp.run(transport="streamable-http")Nice! Now the MCP server is full and may even be deployed regionally:
uv run python most important.pyAny longer, create a GitHub repository and push the native mission listing there.
Step 2
Our MCP server is prepared, however will not be deployed. For deployment, we’re going to use Alpic — a platform that enables us to deploy MCP servers in actually a number of clicks. For that, create an account and check in to Alpic.
Within the menu, select an choice to create a brand new mission. Alpic proposes to import an current Git repository. If you happen to join your GitHub account to Alpic, it is best to have the ability to see the record of obtainable repositories that can be utilized for deployment. Choose the one akin to the MCP server and click on “Import”.
Within the following window, Alpic proposes a number of choices to configure the setting. For our instance, you may go away these choices by default and click on “Deploy”.
After that, Alpic will assemble a Docker container with the imported repository. Relying on the complexity, deployment could take a while. If every little thing goes properly, you will notice the “Deployed” standing with a inexperienced circle close to it.
Underneath the label “Area”, there’s a JSON-RPC deal with of the deployed server. Copy it for now, as we might want to join it within the subsequent step.
Step 3
The MCP server is constructed. Now we have to join it to the LLM supplier in order that we will use it in conversations. In our instance, we are going to use Mistral, however the connection course of must be related for different LLM suppliers.
Within the left menu, choose the “Connectors” choice, which is able to open a brand new window with out there connectors. Connectors allow LLMs to hook up with MCP servers. For instance, in case you add a GitHub connector to Mistral, then within the chat, if wanted, LLM will have the ability to search code in your repositories to offer a solution to a given immediate.
In our case, we wish to import a customized MCP server now we have simply constructed, so we click on on the “Add connector” button.
Within the modal window, navigate to “Customized MVP Connector” and fill within the mandatory data as proven within the screenshot under. For the connector server, use the HTTPS deal with of the deployed MCP server in step 2.
After the connector is added, you may see it within the connectors’ menu:
If you happen to click on on the MCP connector, within the “Capabilities” subwindow, you will notice an inventory of carried out instruments within the MCP server. In our instance, now we have solely carried out a single device “Welcome”, so it’s the solely perform we see right here.
Step 4
Now, return to the chat and click on the “Allow instruments” button, which lets you specify the instruments or MCP servers the LLM is permitted to make use of.
Click on on the checkbox akin to our connector.
Now it’s time to take a look at the connector. We will ask the LLM to make use of the “Welcome” device to greet the consumer. In Mistral chat, if the LLM acknowledges that it wants to make use of an exterior device, a modal window seems, displaying the device identify (“Welcome”) and the arguments it is going to take (identify = “Francisco”).
To substantiate the selection, click on on “Proceed”. In that case, we are going to get a response:
Wonderful! Our MCP server is working appropriately. Equally, we will create extra advanced instruments.
Conclusion
On this article, we introduce MCP as an environment friendly mechanism for creating connectors with LLM distributors. Its simplicity and reusability have made MCP extremely popular these days, permitting builders to scale back the time required to implement LLM plugins.
Moreover, now we have examined a easy instance that demonstrates how one can create an MCP server. In actuality, nothing prevents builders from constructing extra superior MCP functions and leveraging extra performance from LLM suppliers.
For instance, within the case of Mistral, MCP servers can make the most of the performance of Libraries and Paperwork, permitting instruments to take as enter not solely textual content prompts but in addition uploaded recordsdata. These outcomes will be saved within the chat and made persistent throughout completely different conversations.
Sources
All pictures except in any other case famous are by the writer.
