What is MCP (Model Context Protocol)?

Lessons learned 11/11 on Cursor after 15 years in software

We're quite happy when our AI agent writes code or executes simple commands. Less so when we need to manually fetch information in the middle — retrieving API tokens, checking that an element displays correctly, etc.

MCP gives hands to our AI by providing a list of features it can use to interact with a specific program.

Concrete examples

A Google Cloud MCP could for example:

• Launch a new server
• Verify it's deployed
• Check the logs to make sure everything is fine
• Fix your code locally before redeploying

Another one connected to your bank could:

• Search your transactions to categorize them
• Integrate them into your accounting
• Search for invoices in your emails with another MCP

Why it's promising

MCP is probably the most promising thing in applied AI for the real world. Software vendors have understood this well and prioritized its development in their roadmap.

Perhaps too much. We're in the strange situation where many MCPs are written to be first to market. Many of them are buggy, lack documentation, or miss the real problem they're supposed to solve.

Summary

• MCPs will revolutionize AI usage by connecting all useful programs to it
• It's not mature yet, many implementations are botched
• Some are already impressive: Blender MCP for 3D, Chrome DevTools Protocol for debugging, Puppeteer for web automation

What it changes for developers

Until now, AI in code was essentially generation and autocompletion. With MCPs, we move to another level: the AI can act on your environment, not just write text.

It's the difference between an assistant who dictates a recipe to you and an assistant who cooks with you. The first is useful, the second changes everything.

The question is no longer "can AI help me code?" but "what programs do I want to connect to my AI so it can do the work end to end?"

---

Originally published on LinkedIn.