MCP servers - Introduction & Guide

In 2023, Anthropic published the Model Context Protocol (MCP), an open standard often described as a “universal USB-C for AI”. This protocol enables AI systems to connect to external tools, APIs, and data sources, in a bidirectional way.

My take on MCPs:

Any tool can be used by an AI

The MCP standard defines a framework for possible interactions between the tool and the AI. An MCP provides the AI with a catalogue of available tools, from which it only needs to select the right one to solve the given task.

Example: you’re a developer with around twenty tasks listed in Jira. You ask your IDE agent:

You edit the generated action plan, then write:

We’ll mainly focus on MCP server usage by LLM agents (chatbots), though any AI system can use them.

How does a Large Language Model work? The user provides a prompt in text form, which is then tokenised (split into smaller word units called tokens) before being processed by the model. The model is trained on millions or even billions of examples to predict the best possible response, based on the context provided.

Which MCPs to use?

They exist for almost every use case. I mostly use the following:

MCP	Usage
filesystem	Read, create, edit, and search files and directories
sequentialthinking	Facilitates structured, progressive thinking through defined stages.
memory-bank-mcp	Centralized memory-bank with multi-project support
github	Fetch GitHub issues and manage pull requests. → see end of article for integration
google calendar	List your availability for the current week, for example → see end of article for integration
tavily	Search factual information on the web. 1000 free searches per month
fetch	Extract structured data from web APIs
firecrawl	General-purpose web crawler. Free tier offers up to 500 scrapes → see end of article for integration
playwright	Automate tasks in a web browser
shadcn	Integrates Shadcn library components, reducing hallucinations on props

I tested Figma-Context-MCP MCP to convert online mockups into a React application, but the results were not satisfactory.

For those looking to turn a design into code, an alternative approach exists, though it is not particularly recommended:

1. Capture the mockup.
2. Use an LLM to describe the typography, colors, and spacing based on a 12-column grid (desktop), 8-column grid (tablet), and 4-column grid (mobile).
3. Generate the React page along with the associated design system using Shadcn and Tailwind.

Security risks regarding MCPs usage

When you’re running an MCP server locally, it’s the same as running any other software, with unlimited access to all your files.

Instead of running them natively, MCP servers should be containerized and run using Docker with restricted access:

• to the filesystem bu running as a non-root user.
• to the memory & CPU, e.g in docker-compose:

  deploy:
    resources:
      limits:
        cpus: '0.001'
        memory: 50M
          
        

I’ll soon make a guide about using Docker. For the moment, please check out this guide made with strong regard to MCP security: MCP-Checklists

Focus on web MCPs: Fetch vs Firecrawl vs Browserbase vs StageHand vs Playwright

• Need to fetch online site content? → Use Firecrawl. It’s an abstract crawler with rules you can configure. Open-source, cloud-based headless browser you can self-host. It loads sites, executes JavaScript, and retrieves client-side dynamic content, which Fetch cannot do. → No browser automation (no page interaction). → Runs in the cloud and can be integrated into CI/CD pipelines.
• Need to automate actions in a cloud browser? For CI/CD environments, use Browserbase and Stagehand to interact with websites. Difference? Browserbase is imperative: you specify each step. Stagehand is declarative: you specify the outcome, it figures out the steps.
• Need to crawl or interact with a local or authenticated site? Two tools help: Puppeteer and Playwright. Both support headful and headless modes. In headful mode, the browser opens, so you can watch interactions. → Prefer Playwright. It uses the local browser with the current user session, ideal for authenticated sites.

  Playwright sessions can be isolated with the --isolated flag. See Playwright GitHub page.

Setting up MCPs on Cursor

Connect the Firecrawl MCP server

Goal: Crawl anonymous public websites.

1. Go to https://www.firecrawl.dev/
2. Create a Free plan account (up to 500 scrapes).
3. Copy the MCP configuration from the dashboard’s API key section, selecting the latest Cursor integration.
4. Paste it into Cursor settings > MCP > Add new MCP.

Connect the GitHub MCP server

Goal: Access GitHub issues and manage pull requests.

Add this config to your file:

"github": {
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-github"],
  "env": {
    "GITHUB_PERSONAL_ACCESS_TOKEN": "<YOUR_GITHUB_PERSONAL_ACCESS_TOKEN>"
  }
}
        
      

1. Get a Personal Access Token from GitHub
2. Create a new fine-grained token with permissions (all repos or selected ones):
   • Commit statuses: Read and write
   • Contents: Read and write
   • Issues: Read and write
   • Merge queues: Read and write
   • Metadata: Read-only
   • Pull requests: Read and write
   • Secret scanning alerts: Read-only
   • Events: Read-only

Connect the Google Calendar MCP server

Goal: Easily list your availability.

1. Create a new OAuth client ID and secret in Google Cloud Console.
2. Go to APIs & Services > Credentials > Create OAuth ID.
3. Select Desktop app.
4. Download credentials, save as credentials.json at the root of google-calendar-mcp.
5. In Data access, add permissions manually: https://www.googleapis.com/auth/calendar.events

"google-calendar": {
  "command": "node",
  "args": ["<absolute path to /google-calendar-mcp/build/index.js"]
},
        
      

Recap

MCPs are a standard that deliver strong interoperability between AI and external services, to:

• Enhance development environments.
• Enrich workflows by connecting external tools.

Installation is straightforward, and your agent becomes an intelligent hub.

Do you also feel MCPs save you time? If so, which MCP servers do you find most relevant?