The chat window is not the AI
Here's a thing most people never quite untangle: Claude — the model, the thing that thinks — lives in a data center. The claude.ai website is not Claude; it's a website that talks to Claude. Claude Code is not Claude either; it's a terminal program that talks to the same model. Every AI product you've ever touched — the chatbot on a store's website, the "summarize this" button in some app, the AI feature your bank keeps emailing you about — is a thin costume over exactly one move: a program sends text to a model over the internet, and text comes back. The costume is the product.
The talking happens through an API — which sounds technical and is actually just this: a socket in the wall that programs, not people, plug into. People use screens and buttons; programs use the API. Same building, different door.
The meter: what a token is
The API is billed like a utility. The unit is the token — roughly a syllable-ish chunk of text, about three-quarters of a word on average. You pay a tiny amount for every token you send in and every token that comes out, the way a water meter counts liters in both directions. The amounts are genuinely tiny: the model your NPC will run on charges about a dollar per million tokens in, five per million out. A million tokens is roughly ten novels.
Concretely, so you have the number in your head before you ever spend it: one NPC chat message in your game — with everything attached — will cost about a fifth of a cent. The entire course, every conversation and every generated room, lands well under a dollar. The fear is worse than the meter. It almost always is.
The plan: a roommate for the lab
Here's where the six lessons go. Today: the key. Lesson 2: a tiny server — the "brain in a box" that holds your key and talks to Claude on your game's behalf. Lesson 3: the NPC gets a body and its first live words. Lesson 4: a personality, a memory, and — the good part — knowledge of your actual projects, fed from the same lab-data file you built in Course 02. Lesson 5: the NPC stops just talking and starts doing — walking places, changing things — because you'll teach it to answer in data. Lesson 6: a door appears, and every room behind it is invented by a model on the spot.
One architectural note, because it's the professional heart of this course: your game will never talk to Anthropic directly. It talks to a little server of yours, and the server talks to Anthropic. That's not a training-wheels version of the real thing. That is the real thing — the exact shape of every serious AI product on the internet — and lesson 2 is entirely about why.
Get your key
Time to get your plug. The Claude API lives at console.anthropic.com — sign in there (your regular Anthropic account works; the console is just the developer side of it). Three stops:
1. Billing. Find the Billing section and add credit. The minimum is about $5 — that's the toll for this whole road, and most of it will still be sitting there when the course is over.
2. API Keys. Find API Keys and create one. Name it pixel-lab — future-you, staring at a list of keys, will thank present-you for naming things.
3. Copy it somewhere safe — temporarily. The console shows a new key exactly once. Copy it into a password manager, or a private note you'll delete after lesson 2. Not into your pixel-lab folder, not yet — lesson 2 gives it a proper home with a fence around it.
API key created
sk-ant-api03-XXXXXXXX...XXXX
⚠ Copy this now. For security reasons,
you won't be able to see it again.Closed the window before copying? → No harm done. Delete that key, create another. Keys are free — only the credit costs money.
You can now…
- Say what an API is in one sentence — and spot one behind every app you use
- Explain tokens, and why API billing is a meter while your Claude plan is a buffet
- Hold an API key with appropriate paranoia, and know where the revoke button lives
- Name the three-part architecture of every AI product: frontend → your server → model
That vocabulary — API, token, key, "who holds the key?" — is the working language of every AI product meeting on Earth. People are paid handsomely to ask those questions. You now understand all of them, and tomorrow you build the answer.