MCP: Model Context Protocol Server Development in TypeScript

This course contains the use of artificial intelligence.

"I get MCP as a user. But what about building one?" This course answers that question in code.

MCP (Model Context Protocol) is a shared standard connecting AI assistants to outside tools and data. Most material available today leans toward the using side: plugging an existing MCP server into an AI tool. But the real value appears when you move to the other side and expose your own business data and internal systems in a form an AI can call. This course covers only that building side, as 50 exercises in TypeScript.

This is not a course you read, or a course you watch. All 50 exercises run to the same rhythm: task, acceptance criteria, implement it yourself, grade it with npm run check, then check your work in the review video. There are two videos per exercise. The problem video hands you the knowledge and the expected result, you pause there and write it yourself, and the review video covers the model answer, common errors and the design decisions. Nothing about this course lets you watch it through with your hands still.

The curriculum is six chapters. Chapter 1 covers Tools in 10 exercises: registerTool, Zod schemas, structured output, error design, progress notifications and logging. Chapter 2 covers Resources and Prompts in 8 exercises: URI templates, binary data, subscription notifications, prompts with arguments, and resource_link. Unlike Tools-heavy material, all three primitives are practised evenly, by hand.

From Chapter 3 the servers get practical. In the file exercises, the test actually attacks your server with path traversal. Read-only SQLite queries, REST API wrappers, rate limiting, CSV analysis, pagination. It also gets into a design question specific to MCP: how to return results without eating up the context window. Chapter 4 turns to the client side. In exercise 29 you write the test harness that has been grading you all along, which pulls back the curtain on the grading system. You then implement sampling, elicitation, roots and completion, and finish by assembling a mini agent that runs a tool use loop, so you feel in your hands that MCP is a component inside an agent.

Chapter 5 is going remote and authentication. You move from stdio to Streamable HTTP, design it stateless (and prove with a test that it still works when requests are round-robined across two independent servers), validate Origin, implement Protected Resource Metadata (RFC 9728), and hand-write JWT verification with no external library so that tools are exposed by audience and scope. Chapter 6 is quality and extension. Exercise 44 is the one with the structure reversed: the server is finished, and you write the tests. Prompt injection defence, tracing with OpenTelemetry, the Tasks extension, the groundwork for MCP Apps, and publishing to npm.

The last exercise, 50, is a capstone. You turn a problem from your own work into an MCP server, write ten realistic questions an LLM ought to be able to answer against it, and verify with a real client. Your tests turning green and an LLM genuinely using your server well are two different problems, and experiencing that gap is how the course closes.

No local Node installation is required. Upload the exercise ZIP to Google Cloud Shell and run the included setup script to prepare the dependencies. Node 24 or newer supports all 50 grading exercises. A Google account is required, with a free quota of 50 hours per week and 5GB of storage. Cloud Shell is unavailable to accounts under age 18 or with an unknown age. A local Node 24+ alternative is included. When videos show CodeSandbox, use the updated startup lecture instead. Grading requires no API keys or paid Google Cloud resources; external API tests use stubs.