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.
Only cards that are actually free · owned − committed
Build and Upgrade draw only from cards you own — and if you lock a deck (mark it
as one you've physically sleeved), its cards are treated as spoken-for and left out of your other
builds. It's quantity-aware: owning two copies of a card lets two locked decks each keep one,
and Manaforge flags any card committed to more locked decks than you own copies of. So a new
deck is assembled from what's genuinely available, not from cards already in a deck on your table.
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. Upgrade pursues combos harder
at higher brackets: at Bracket 3 it finishes combos you're a single owned card away from; at
Bracket 4–5 it also assembles multi-card combos you own the pieces for, buys the missing pieces
(within your budget), and ranks combo completion ahead of ordinary synergy swaps. The Combos
control on the Upgrade screen lets you override this — combo-heavy to pack in as many as
possible at any bracket, or fewer combos to keep only what's already live.
Tuning it yourself · you stay in control
The defaults are calibrated per bracket, but nothing is a black box — two panels on the
Build and Upgrade screens let you drive the model directly (they start collapsed; open them any time):
- Fine-tune the scoring — sliders for the four weights above (synergy · popularity · curve ·
efficiency). They re-rank which cards make the cut without changing the deck's shape;
normalised to 100%, and identical weights always yield the identical deck.
- Reshape the deck — role targets — set exactly how many lands, ramp, card draw, removal,
board wipes and protection to aim for. This changes the deck's structure, which the weights
alone can't. A live budget readout shows how many nonland slots the targets consume.
Touch nothing and the engine's own bracket defaults apply — so tuning is opt-in and
never drifts your results unless you ask it to. Each panel resets to its bracket default in one click.
How it was built · full transparency
Honest disclosure: I'm a researcher and a Commander player, not a professional developer — I
"vibe-coded" Manaforge with heavy help from an AI coding assistant. I designed it, made the
decisions, and tested it against real decks; the AI wrote a lot of the code alongside me. I'm telling you
because you deserve to know. The important part: that's build-time only — no AI runs when you use the
app. Everything that happens when you build, analyse, or upgrade a deck is the deterministic,
inspectable maths described above. AI helped make it; it doesn't make your decks.
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; the AI that helped write the app never 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.
