The problem

Most deckbuilding tools are autocomplete — they suggest the next card given a partial list, with no regard for what's in your binder. The output is almost always a card you don't own and would have to buy. That's fine for tournament netdecking; it's not useful if your goal is to play decks built from your existing collection.

Inventory-constrained deck construction flips this. Every nonbasic card the builder uses must be something the collection actually contains; basic lands are added automatically. The output is a deck you can sleeve up tonight.

How decks get picked

A beam-search engine assembles candidate decks under the rules of the selected format (legality, singleton, color identity, commander eligibility, etc.). Learned rankers help order card choices in every format; where supported, a small neural scorer compares completed candidates. These models run on the builder's own server. They do not invent cards or override the collection and format constraints. The top survivors are refined through a constraint-bounded exchange step that swaps weak slots for stronger ones from the same pool.

Who's behind it

Built and maintained by the mtgsold team. EDHREC-derived deck data is used with permission from EDHREC; card data comes from Scryfall. The project is fan-made and not affiliated with Wizards of the Coast.

What it isn't

• Not a card-buying service. No purchases are offered, brokered, or guaranteed.
• Not a Wizards product. Not affiliated with, endorsed, or licensed by Wizards of the Coast.
• Not a tournament authority. Format legality and card availability should always be verified independently before sanctioned play.
• Not a replacement for taste. The builder's rankings are useful but imperfect; treat its top result as a starting point, not a final answer.