MCP Tutorial: Build Your First MCP Server

It shows the whole loop in under 14 minutes. Scaffold, write functions, install into a client, restart, ask a question, watch the tool call. For someone who has only read about MCP, seeing Claude Desktop list the three tools and pick one from a vague question makes the protocol concrete.

Docstrings as the contract is exactly right. The model never sees your code, only the tool name, parameter schema and description. The date-format example proves it. When a tool misbehaves in practice, the description is usually the first thing to fix.

The mcp install shortcut saves a common failure. Hand-editing claude_desktop_config.json with the wrong path or interpreter is where many first attempts break. Showing the generated entry afterwards, rather than hiding it, helps viewers debug later.

State does not survive a restart. The data lives in a Python dictionary, so the July 4th booking disappears when Claude Desktop relaunches the server. Swapping in SQLite or a real HR database is the obvious next step and is not shown.

The write tool has no guardrails. Apply leave does not check that the dates are in the future, that the balance is sufficient, or that the request is not a duplicate. Nothing checks who is asking. A model can be talked into calling write tools by text in a document or chat, so the server, not the model, must enforce rules. Claude Desktop's approval prompt at 12:55 helps, but it is a client feature, not your protection.

Resources and prompts are mentioned, not explained. The greeting resource is never used in the demo, so viewers do not learn when data belongs in a resource (something the client reads) rather than a tool (something the model calls).

Local stdio only. This server runs as a child process of Claude Desktop on one machine. To reach it from a web chat app or a teammate's client, it would need to run as a remote server over HTTP with authentication, which, as of the video's April 2025 upload, was still settling in the spec.

Some setup is redundant. Installing the SDK globally with pip and then managing the project with uv mixes two environments, which is a common source of errors like the typer issue he hit.

Swap leave balances for your own notes and the same pattern becomes a memory server: a search tool and maybe a save tool, with docstrings that tell the model when to use them. The gotchas carry over too. Persistence, sensible limits on what a write tool may change, and remote access are what turn a demo into something you rely on. The guide to giving your AI memory over MCP covers those steps.

If you would rather not run that server yourself, BrainTube is a hosted option for research material: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. For other approaches, see MCP memory servers compared.