Manaforge is a decision-support system, not an autopilot. It gathers the
same evidence a strong deckbuilder would, scores every option against explicit, measurable
criteria, and shows you its reasoning. There is no black-box AI making the choices: the same
inputs always produce the same deck, and every card can tell you why it earned its slot.
The evidence
Three authoritative, openly published sources — all queried live from your own
browser, so your collection never leaves your device:
Scryfall
Ground truth for every card: mana value, type, rules
text, colour identity, format legality, market price.
EDHREC
Aggregated statistics from hundreds of thousands of real
Commander decklists — the wisdom-of-crowds signal.
Commander Spellbook
A human-curated catalogue of thousands of
known card combos and the effects they produce.
The decision model · multi-criteria decision analysis
Every candidate card is scored with a weighted-sum model — the classic method
of multi-criteria decision analysis (MCDA) — across four normalised criteria:
score = wS·Synergy + wI·Inclusion + wC·CurveFit + wE·Efficiency
- Synergy — how much more often a card appears with your commander than at random, from EDHREC.
- Inclusion — its empirical inclusion rate: of the decks that could run it, how many do (decks ÷ potential decks).
- Curve-fit — how well its mana value fills the gaps still open in a healthy curve, recomputed after every pick.
- Efficiency — a mana-cost term that favours cards doing more for less.
Criteria are min-max normalised so none can dominate; crowd-favourite and
power-level adjustments are then layered on. The weights shift with the bracket you pick —
a casual build leans on synergy, a cEDH build on efficiency and raw inclusion.
Assembling the 100 · constrained optimisation
Scoring ranks the cards; a constrained greedy optimiser builds the deck. It
fills role quotas in priority order — ramp, removal, card draw, wipes, protection — under the hard
constraints a legal, functional Commander deck must satisfy:
- exactly 99 + commander, singleton, inside the commander's colour identity;
- a minimum creature count derived from the archetype's own creature ratio;
- a curve-trim pass that swaps overweight top-end for on-curve cards;
- a colour-balanced land base, apportioned to the deck's actual coloured-mana demand.
Then the official Commander Brackets are enforced as measurable caps — how many
Game Changers, tutors, mass-land-denial pieces, extra-turn spells and two-card combos the target
bracket permits. Upgrade can raise or lower a deck to fit its bracket.
Measuring speed & power
Instead of guessing, Manaforge computes a transparent 0–100 speed index — a
linear combination of measurable acceleration signals (fast-mana density, cheap ramp, low curve,
tutor count) calibrated against real decks of known power. Honest limit: this is a density
proxy, not a turn-by-turn simulation — it estimates a deck's potential to go fast, not the
exact turn it wins.
Combo detection · a graph model
Your list is matched against Spellbook's catalogue as a graph: each combo is a
small network of cards linked to the effect they produce. Every combo is classified live
(all pieces present), one card away (a single addition completes it), or unreachable —
which is exactly what drives the combo map and the buy suggestions.
Why you can trust it
- Deterministic — the same inputs produce the same deck, every time.
- Explainable — every pick shows its role and reason; nothing is hidden in a model.
- Transparent weights — the criteria are constants you can read and tune, not learned parameters.
- No runtime AI — the maths above is all that runs when you build. An LLM helped write the app; none runs it.
- Honest about limits — crowd data reflects popularity, not correctness; a greedy optimiser is fast and near-optimal, not provably global; speed is a proxy. We flag these rather than bury them.
