Frequently asked questions

No. The site signs you in anonymously in the background so your game has somewhere to live. You never enter an email or a password, and there's nothing to log into.

Your game is saved against that anonymous identity, so closing the tab and coming back later on the same browser picks up where you left off. Clearing your browser storage loses the link to it.

Open the deck menu in the top bar and paste a decklist. SwooLab reads SWUDB's "Copy JSON" export (which Karabast and most SWU tools also use), melee.gg's "Copy to Clipboard" text, and the SWU.com text format. To build or adjust a deck, open the deck editor; decks you save go in My Decks.

You can set a deck for either side, and swap which one you're playing. A new deck takes effect on the next game, not mid-game. Start a new game after importing.

In your browser's local storage, on this device only. They don't follow you to another browser or device, and clearing this site's data deletes them.

Sealed is a simulated sealed event. Open six packs and build a 30-card deck from them, with the two pre-release kit leaders always available. Then play three Swiss rounds against seven AI opponents.

The opponents' decks come from a library of 1,000 decks built by simulating a full sealed event with the AI (how it was done).

Thinking time is how many positions the AI's search is allowed to look at before it commits to a move. None skips the search entirely and plays a move using the policy network, which is fast and noticeably weaker. Low, Medium, and High each roughly double the search.

The AI can't see your cards. Instead, during search, it samples plausible versions of the game consistent with what it has seen and averages over each one.

You can step backward and forward through the whole game with the floating arrows or the arrow keys, and you can jump to any point by tapping an entry in the action log. Making a move from an earlier position continues the game from there and discards what came after.

The model's own estimate of your chance of winning from the current position. It's the value the network learned during training, not a simulation. Treat it as a strong opinion rather than a measurement, especially in unusual board states.

Every released set, from Spark of Rebellion through Homeworlds.

If a deck includes a card that hasn't been implemented yet, you'll get a notification.

Please report it. The bug-report button in the top bar attaches a compact replay of the game, so the position can be reconstructed.

A note about what you expected to happen makes it much faster to track down. Links or references to any relevant rules or judge rulings are also appreciated.

Strong enough to punish loose play and to find lines that aren't obvious, and still beatable. Its strength also depends on the settings: at None thinking time it plays a fast, shallow game; at High it plays considerably better.

It is weakest in positions unlike the ones it trained on, which in practice means decks and card combinations it rarely saw.

No. There is no human game data anywhere in the training. It started from a randomly initialized network that knew nothing but the legal moves, and improved by playing against itself.

The decks it plays are another matter. Early training is done with uniform random decks using simple heuristics (like limiting the number of out-of-aspect cards). In the later stages of training, decks were drawn from a mixture of random decks, decks built by a heuristic that favors cards with synergies, and decklists reported from tournaments.

The long term goal of this project is to be able to learn without any human data at all, but the deck models that will make this possible aren't ready yet.

The search runs on a server that scales to zero when nobody is playing, so the first request of a session may have to wait for it to start up. Subsequent turns should be much faster. This can also happen if the server is busy.

The AI model is trained to play the best it can, but it does still make mistakes. This is most noticeable when using None thinking time, which is fast but doesn't look into the future at all. This can lead to the AI doing dumb things.

No. SwooLab is an unofficial fan project with no connection to Fantasy Flight Games or Lucasfilm. Card names and images belong to their respective owners.

Yes. The site has a dedicated mobile layout. Tap a card to see it full size.