MCP Tutorial: Build Your First MCP Server and Client from Scratch (Free Labs)

Coverage is the strength. Most beginner MCP videos stop at tools. This one covers all six primitives, both transports, the JSON-RPC envelope, the Inspector, and the context object for progress updates. The section on sampling explains the design reason, not just the API: the server stays decoupled from any model, and the client keeps control of model choice and cost. That reasoning is what makes MCP servers reusable across clients.

The warning at 22:25 is also correct and easy to miss. A remote server receives your data, so authentication, authorization, privacy and trust in the operator matter more than they do for a local stdio process. The same goes for community servers, which he says at 9:59 come with no guarantee of testing or currency.

The lab flow is a sensible order: use a server, then build one, then build a client. Seeing a tool approval dialog in Roo Code before writing any code teaches more about the user experience than a diagram does.

• How tool choice really works. He calls MCP a guide that helps the AI choose the right APIs, but never shows that the tool name and description are what the model reads. In practice the quality of the description and the parameter schema decides whether the model calls the tool correctly. A lab step that rewrites a vague description and watches behavior change would teach more than the multiple-choice questions.
• Who controls each primitive. The spec's design is that tools are model-controlled, resources are application-controlled and prompts are user-controlled, usually surfaced as commands or menu items. The video presents prompts as a better system prompt the server author writes for the agent developer. That is partly true, but it hides that a user normally picks a prompt, and that many clients pass resources only when the user or app attaches them.
• Transport history. He lists stdio and HTTP and mentions streamable HTTP. As of the July 2025 upload, streamable HTTP had replaced the older HTTP plus server-sent events transport in the spec, and older tutorials still show the old one. Check which one your client supports.
• Security beyond one sentence. Tool descriptions enter the model's context, so a hostile server can inject instructions. Roots limit file access only if the server respects them. Remote servers in the spec use OAuth-based authorization, which the video does not show.
• The agent section is long for a build tutorial. Roughly the first ten minutes are agent background, and the slide at 8:36 about debugging a UI regression through git history is described as not using an MCP server, which muddies the use-case list.

The video's resources section maps directly onto personal knowledge. Airport lists and refund policies are the same kind of thing as your notes and saved transcripts: reference data an agent should consult before it acts. The design choice that matters is how you expose it. A large knowledge base works better as a search tool than as a list of resources, because the model can query it on demand instead of the app loading everything up front. That is the pattern behind an MCP memory server.

BrainTube follows it. It saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes that as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search the same material. If you want to build your own instead, the FastMCP recipe in this video is enough to start. Expose one well-described search tool first, then add resources only for small, stable reference files. For a setup walkthrough see give your AI memory with MCP.