How MCP Connects AI to External Tools

▶ Watch (00:11)

MCP, Model Context Protocol, is an open standard that lets AI models connect to tools, APIs, and data sources through one consistent interface. VS Code is the host. GitHub Copilot is the client. Lightweight MCP servers are the programs that expose specific capabilities to both. Adolphe’s shorthand for the whole thing:

“it’s kind of like a USB-C, but for AI” — Reynald Adolphe

The payoff is AI that pulls from external systems and gives context-aware answers without exposing private data.

Publishing a Repo Without Touching the Terminal

▶ Watch (02:11)

The GitHub MCP Server connects Copilot directly to GitHub’s platform, giving agent mode the ability to read repositories, manage issues, and push code. Adolphe installs it from the VS Code extension marketplace in seconds, then asks Copilot to initialize and publish a budgeting app. He walked through the GitHub agent mode repo push (04:22) live: agent mode ran every git command in the background, and when a push failed, it retried automatically.

“if it screws up, it keeps working at it until it gets it right” — Reynald Adolphe

The repo appeared on GitHub complete with a README describing the tech stack and setup steps.

Grounding Code Generation in Official Docs

▶ Watch (05:54)

The MS Learn MCP Server pulls trusted, up-to-date Microsoft documentation directly into Copilot, so generated code follows the patterns on learn.microsoft.com rather than whatever the model was trained on. Adolphe asked it to build a CRUD app using Entity Framework and explicitly told it to use the Learn docs. Copilot stayed focused, cited the specific tutorial pages it referenced, and built a running MVC app at port 5000.

“I didn’t waste time looking at all these other tutorials and documentation” — Reynald Adolphe

The app matched the learn.microsoft.com tutorial layout and was fully functional on first run.

Generating Browser Tests from a URL

▶ Watch (09:25)

Playwright MCP Server handles automated browser testing from a URL alone, no app source code required. Adolphe’s project contained only a prompt file that instructed Copilot to create a test file when finished. He watched the Playwright IMDb live navigation (10:51) in real time: Playwright opened IMDb, found the search bar through the ads and pop-ups, typed “Heat”, selected the correct film, and then generated a complete test file in a new folder. A site with 100 pages now gets its test suite written automatically rather than by hand.

Notable Quotes

it’s kind of like a USB-C, but for AI Reynald Adolphe · ▶ 1:05

if it screws up, it keeps working at it until it gets it right Reynald Adolphe · ▶ 4:43

I didn’t waste time looking at all these other tutorials and documentation Reynald Adolphe · ▶ 8:51

Key Takeaways

  • MCP servers let Copilot agent mode push code to GitHub without any terminal commands.
  • The MS Learn MCP Server pins generation to official Microsoft docs, so the output follows current best practices.
  • Playwright MCP Server browses a live site and writes a working test file from a single prompt.