Engineering courses · AI infrastructure

The production layer of AI, taught from the spec up.

Courses for engineers who have to keep the thing running. Each one starts from failures you'd get paged for, and every lab runs in Docker on your own machine. If a lesson says the server returns -32020, there's a test in the lab that makes it.

See MCP Production Read the migration guide

lab-14-reference — two replicas, one SQLite file

$ python3 conformance.py
PASSAn Origin that is not allowed gets 403, with no id in the body
PASSA request id is never null
PASSMcp-Method that disagrees with the body gets 400 and -32020
PASSMcp-Name that disagrees with the body gets 400 and -32020
PASSAn unsupported version gets -32022 listing the supported ones
PASSserver/discover answers on the first request, with no initialize
PASSDiscovery and list results carry ttlMs and cacheScope
PASSEvery result names its resultType
PASSAn unknown tool is a JSON-RPC error (-32602), not a tool result
PASSAn unknown method gets -32601
PASSA request without a token gets 401 naming the resource metadata
PASSProtected resource metadata names this resource
PASSA token for another resource gets 401
PASSA notification gets 202 with no body, or an HTTP error if refused
… 8 more rows
OK  14 of 14 rows hold
$ 

Real output from the reference server in MCP Production, lab 14.

01 / How a lesson gets made

Nothing gets said on camera that a test hasn't already shown.

Here's one fact from the MCP course, followed from the spec to the lab to the narration, with the check that keeps it honest afterwards.

Claim1

Written down with its source

Each fact a lesson relies on gets a record: the sentence, the spec section, the date it was checked.

1.3-mismatch
Required header validation failures return HTTP 400 with HeaderMismatch code -32020.
checked 2026-09-09

Lab2

Proved by a test

The lab sends a request whose header disagrees with its body, and checks what comes back.

PASS Mcp-Method that disagrees with the body gets 400 and -32020

Lesson3

Then the narration

The script is written against the passing lab, and the terminal you see in the video is that lab running.

Lesson 1.3 · Diagnose a rejected call without guessing at the server

Gate4

Checked again later

Gates rerun across every lab. One caught a gateway answering refusals with -32000, a code the revision tells new servers not to use. Every test still passed. The gate didn't.

qa_errorcodes.py
lab code uses only the error codes its revision allows

02 / Courses

One course out now. More on the same standard.

Model Context Protocol · revision 2026-07-28

MCP Production: secure, scalable MCP servers in Python

Build MCP servers in Python that hold up behind a gateway, across replicas and under real OAuth. Then break them on purpose and watch the tests catch it.

57 lessons, 6h58 of video
14 local Docker labs
Python, with TypeScript in module 1
No cloud account or API key

Free · PDF

Moving a server from 2025-11-25 to 2026-07-28?

The cheatsheet puts every assumption the new revision breaks on the left and what replaces it on the right. Error codes, the deprecation clock and a staged rollout checklist are on the pages after.

03 / Who's behind it

One engineer, held to the tests.

Runs Clean started with a software engineer who kept seeing the same gap: courses end where production starts. The demo works, then the second replica, the retry and the expired token arrive, and nobody taught those.

So every course here starts from the failure. Claims are sourced and dated, labs pass before a lesson is recorded, and each course names the spec revision it covers. You don't have to take any of that on trust, because every lab ships with its tests.