Course catalog
Ship an MCP server and a skill that a real client actually loads.
Connect AI to Tools (MCP) taught you how the protocol works. This course is about shipping your own server and skill. You start from the caller you are really designing for — a model that never opens your docs, routes on your tool names, and pays tokens for every word you return. You choose between tools, resources, prompts, and skills; trace a session from initialize through tools/list to tools/call over stdio and streamable HTTP; and budget what your tool list costs on every single turn. Then you design the parts a model touches: names and descriptions, schemas that reject the bad call before your handler runs, return values small enough to act on, and errors that say exactly what to change. You scaffold a stdio server, register it with a real client and watch it go from pending approval to connected, and author a SKILL.md whose frontmatter actually triggers. You finish with the trust work — evals that separate routing from execution, the four failure modes you meet in production, permissions designed for consent, and versioning and publishing something other people can audit before they load it.
Section 1
What a model actually sees of your server: primitives, transport, and the token cost of your tool list.
Section 2
Names, schemas, return values, and errors — the four surfaces a model actually touches.
Build and run a real AI product: it searches your own content, uses only the tools you approved, is tested the same way every time, has a plan for when it fails, and shows its speed and cost.
11 course sequence
Ship a secure TypeScript app built with AI in your workflow: tested algorithms, tools kept inside safe limits, architecture choices you can explain, and releases you can check after they go out.
16 course sequence
7-day free trial on every plan for first-time members · Cancel anytime
Get startedSection 3
Scaffold a stdio server, register it with a real client, and author a skill the agent loads on its own.
Section 4
Evals that separate routing from execution, the failure modes you will actually meet, consent by design, and shipping something people can audit.