ManaBox exports are the supported format right now (the file ManaBox produces from Export Collection). The parser reads quantities, set codes, conditions, languages, foils, and purchase prices straight from the standard ManaBox columns. Support for other trackers (Moxfield, Archidekt, TCGplayer, etc.) isn't built in yet, though more may be added over time.

Pauper, Commander, Legacy, Brawl, Jumpstart, Paupercommander, and TinyLeaders. Each format has its own legality rules, deck size, color identity logic (where applicable), and a learned card-ranking model trained on real decklists. Six formats also use a calibrated neural deck scorer; Jumpstart uses the structural composite score at the deck level.

It uses machine learning, but it is not an LLM or generative-AI service. Models running on the builder's own server help order candidate cards and compare completed deck candidates; no third-party AI API is involved. The engine never invents cards: nonbasic choices come from your collection, with basic lands supplied by the mana-base allocator. The models cannot bypass inventory, legality, deck-size, color-identity, or other hard constraints, and uploaded collections are not used for model training.

If the card isn't legal in the format you selected, the picker shows up to 5 near-name matches and tells you the formats the card IS legal in. If the card is from a set that hasn't released yet, the picker flags it as Unreleased with the release date — Scryfall stamps not-legal on every format until the set drops, so the picker translates that into a useful message.

The grade summarizes structural measurements such as synergy, curve fit, completeness, castability, and win-condition coverage. It is not a win-rate estimate, power-level guarantee, or tournament-readiness certification. Use it to compare this builder's results within a format; you can filter out lower grades in the Advanced section of the build preferences.

The engine assembles candidate decks with beam search under hard format and inventory constraints. A learned ranker helps order card choices; in supported formats, a small neural scorer compares completed candidates. The models do not invent cards or override the constraints. A final exchange pass may replace weak slots with stronger choices from the same collection pool.

Every nonbasic card the builder considers must be present in the CSV you uploaded, with enough copies for whatever slot it fills. Basic lands are the one deliberate exception: the mana-base allocator adds them automatically. The output does not suggest that you buy four Force of Wills you don't own.

Your own CSV. The builder treats whatever price column you uploaded as ground truth and reports per-deck totals using those values. No purchase is offered or brokered through this site.

They're added automatically by the builder. After the spells and nonland slots are filled, a mana-base allocator computes how many of each basic (Forest, Island, Mountain, Plains, Swamp — Wastes for colorless) the deck needs based on its color identity and curve, then injects those copies into the decklist regardless of what your collection contains. Two consequences: (1) you don't need to own any basics in your uploaded CSV for a deck to build successfully; (2) the Download picklist and Download remaining CSVs exclude basics, because they were never drawn from your collection in the first place. Bring your own basics — most fat packs and prerelease kits supply them in abundance.

It shouldn't be. If you see a sub-target deck, the build hit a wall during refinement — usually because the collection didn't have enough copies of the higher-scoring swap candidate. Re-run the build; if it persists, it's worth reporting with the CSV so it can be looked into.

Yes — Step 2 has a "Lock specific cards" section. Pick a legal commander for commander-style formats; pick any legal nonland card as a required include. If your collection doesn't own the locked card the build fails fast and tells you so.

Builds usually finish within a few minutes. Commander builds with a small commander pool take longer; pauper or jumpstart builds with a large CSV are quicker. The progress panel updates live so you can watch what happens search round by search round.

Uploaded CSVs are stored temporarily in Cloudflare R2 and automatically deleted after three days. Saved Crew inventory remains in your account until you change or remove it. Neither uploaded CSVs nor saved inventory are sold or used to train the ranking models.

Cloudflare Turnstile gates the build endpoint so automated abuse can't drain the compute budget. It's a one-click check (no puzzles), and only the build action is gated — browsing and searching are not.