The Illustrated Guide to Local AI

Build Your AI Lab

A board game for two to four players. Build a lab out of all six decks — hardware, models, tools, advisors and partners. Learnable in five minutes; the Starter Lab is the first-timer path.

The pitch

You and the other players are running labs. You build an AI lab by playing cards from every deck: you buy the hardware, you get the model, you write the code, you bring in the builders, you partner with the companies. The first player to finish a Flagship Run — getting a model to work end to end on their own lab — wins.

The five resources

Every card costs some of these. Pay from your bank.

ResourceWhat it buys
ComputeProcessor and graphics cycles — running things.
VRAMModel memory — fitting things in.
DataTraining and fine-tuning fuel.
TalentPeople time — setup, fixes, research.
PowerElectricity — the meter that never sleeps.

Two modes

The rule that makes it a teaching game. Open-weights models can be played in both modes. Closed models can only be played in Cloud Lab. That single rule is the real-world choice the guide is about — and in the game it costs you something.

1. Setup

2. A turn (three steps)

3. Building your lab (the card roles)

Each type of card has a job in the lab. The schema's card types map to these roles:

TypeRole in the labHow it works
HardwareFoundationPlayed first, in the Foundation row. A model can only be played if the Foundation can hold it. Foundation cards grant standing bonuses.
ConceptEngine and upgradesYou may have exactly one model card in play at a time. Concept and technique cards modify the model's costs and stats.
CommandTools (operations)Played to do things: move a model file, check your totals, stop a failing Run. Each command can be played once per game unless it says otherwise.
CodeTools (code)Played to fix or build: add a bonus to a Run attempt, ignore one failure, or pull a partner or technique from the supply.
PersonAdvisorsNot played like ordinary cards. Put a creator in the Advisor row by paying their Talent cost once. Mythic advisors stay for the whole game. Jokers may instead be played as Tools, for their printed effect.
CompanyPartners and archetypesA company card is a standing partner: it grants its signature bonus while in play. Playing a closed company — or a closed model — in Local Lab is illegal. That is the whole point.

Every card lists its own cost. Pay from your bank. In Cloud mode, every Compute and VRAM you spend also costs one Credit.

4. Advisors, origins and combos

5. Running a model (the win check)

To attempt a Run, your lab must pass three checks. Nothing is rolled — the checks are arithmetic, which keeps the game beginner-safe.

Success is a Run. A Flagship Run — a success with a powerful model — wins the game for you. Failure means the model crashes: you lose one Power, and the failed card goes to the discard. A command tool such as kill can be played to abort the failure without losing the Power.

6. Jokers — the harmless chaos

Jokers are the playful creators: famous, loud, and on purpose weak. They are played once, as Tools, for their printed effect. The tone rule for every joker is simple: the joke is about the public role only — the hype, the rebrands, the announcements. Nothing personal, nothing cruel.

7. Winning

The Starter Lab

The Starter Lab is a small, fixed box for a first game. It holds just enough cards to build one working lab: a little hardware, a single model, two or three tools, and one advisor to choose from. Everyone starts from the same friendly pile, so the first game teaches the shape of the thing — hardware, then model, then tools — without the walls of a full collection.

Play one Starter Lab game in Local mode. The win should feel like a small, honest victory: a model that fits, tools that work, and a run that runs. Add the full collection afterwards, and the walls come with it — VRAM pressure, closed models, and the Cloud bill.

How new cards join the game

The game never hardcodes a card. It reads every card straight from the normalised data files and looks at its rarity, type and stats. Because of that, a new card is playable the moment it is added to a deck: no rule rewrites, no special cases. For the exact steps, see ADDING-CONTENT.md in the project root.

First-timer checklist