Data security and popularity of the tools are to be taken into account.
Until then, despite all its power, AI could not know what was on your personal computer, the specific details on your business or your documents. But with the MCP, for Model Context Protocol, the different models of AI can now connect to various data and tools in a standardized manner.
The MCP operates on the basic principle of the web: a customer-server architecture. The host is the application that communicates with AI models, like Claude. The MCP client, such as Claude Desktop, or Cursor, makes it possible to collect and issue requests, in JSON-RPC 2. The MCP server, which can be deployed locally or in the cloud, allows access to external tools and sources.
Easy to use tools
In order to help you choose from the number of MCP booming MCP servers, we select some. Currently, trends concerning them revolve around a few areas such as integration into current productivity or collaboration tools. We think of platforms like Google Workspace and Microsoft Office 365. There are also a number of software development tools, such as Github and Gitlab, developing MCP servers. They manage deposits, or even analyze merger queries. Another interest: interaction and web research. Brave Search for example allows you to navigate the sites and extract data.
To choose them, we trust the names of companies like Playwright, and/or the number of stars of the server on Github, or the quality of the site delivering MCPs. This should allow, in part, to avoid risks such as data leakage or the proper functioning of the LLM. We are also interested in MCP servers relatively easy to use and the tools allowing, for the most part, relatively simple actions, in order to speak to as many people as possible.
Cursor as MCP customer
Installing MCP servers locally on the machine often requires monitoring the indications on Github. This requires copying the appropriate source code, installing outbuildings and setting up the additional elements.
In our case, we chose Cursor as MCP customer. We collect the JSON configuration file from GitHub for example, or via the Cursor site. It is then integrated into a .cursor folder and an MCP.json file. Note that we have previously configured the AI -based editor in the “Settings” section. Objective, in particular: or not to decide on the participation of models like Claude 4 SONNET or GPT-4.1. Let us indicate that prompts are carried out via a cat.
1. Filesystem
The MCP Filesystem server allows you to read, get metadata or write files. It also serves to create, list, move, search or delete directories. To set up it, he requests the specification of a directory in the JSON installation file. Note that it uses the NPX command (Node Package Execute). It allows you to carry out NODE.JS packages directly, without installing them overall on your machine or locally in your project. In this case, it gives:
"filesystem": {
"command": "npx",
"args": (
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/bruno/bureau",
"/path/to/other/allowed/dir"
)
}
Note that the “MCP.JSON” file in question must start with: “McPServers”: {
This MCP server is therefore relatively easy to install.
For example, we can ask him to create a folder with a file. Just mention this in simple language in the cat. Our tests are conclusive. The latency period is quite short.
2. Playwright
The MCP Playwright server interacts with web pages, makes screenshots, generates test code, or even performs JavaScript in a real navigation environment. We have implemented it easily, including the tools offered by Cursor. The code is this:
"Playwright": {
"command": "npx @playwright/mcp@latest",
"env": {}
},
We asked for it to make a screenshot of a web page. After a few calculations, it offers it, in the specified path, with the name of the file. The page was copied in its entirety.
(Playwright-Server-MCP)
The ease of installation of this server and its versatility are appreciable.
3. Browsebase
The Browsenbase server is capable of extracting structured web data, or navigating, clicking and identifying the forms to be completed. Its installation requests to enter an API key and a project_id. For example, we asked him to identify the forms to be completed for a particular website:
The result is satisfactory. The screen copy function has proven to be less conclusive. This type of server can be interesting, especially from a SEO perspective, with the management of structured data.
4. Stripe
The MCP Stripe server connects to the Stripe payments processing services. It allows you to create a new customer, product, price, payment link, update an existing invoice, or even create a new coupon. Its installation requires the “secret key” Stripe, to be recovered from the site in question.
We asked the MCP server to create a client profile with the associated names and email. After a short period of latency, the information requested appear well on the site intended for internet payment.
This server must thus make operations concerning the integration of payments less time -consuming.
5. GitHub
The GitHub MCP server automates workflows and processes, or even extracting and analyzing data from GitHub standards. On the installation JSON file, the “Docker Run -i –RM -E GitHub_personal_Access_Token ghcr.io/github/github-mithb-mc-erver” command means, in simple terms: download and launches a portable and temporary application of GitHub, giving it a password so that it can work. We ask the server to correct code:
He runs and gives simple explanations. Developers who use AI tools to code appear as a target of choice for this server style.
