EDH Seer

Paste a Commander decklist and see which of your cards actually work together, with the sentence that says why. Free, no account, nothing stored. Paste a list, or a Moxfield or Archidekt link, at edhseer.cards.

What you get

• Every pairing in your deck, in a sentence you can check. Not a score and not "people also play". A claim that names both cards and the mechanism:

  "When Siege-Gang Commander dies, Skullclamp draws you 2 cards"

  Open both cards. If the printed text does not say this, the claim is wrong, and a wrong claim is a bug you can report with two card names.

• What your commander works with. Your commander in the middle and every card it works with around it, grouped by what the link is. Tap a card to read how, or put any card in the middle and follow its links from there.

• What to fix, biggest payoff first. Card draw, interaction and answers counted against benchmarks, with the gap stated in cards: "You are 2 short on interaction. Add ~2; most cost 2–4 mana."

• Mana you can plan around. Land count, colour sources, and the odds of casting each of your hardest spells on curve, from 2,000 simulated games.

• Your game plan. The themes your deck leans on, and the pairs behind each one.

• Bracket, roles and combos. Where the deck sits, what each card is doing, and the combos it already holds, from Commander Spellbook.

What your deck does	What to fix, biggest payoff first

No deck handy?

Every commander, every card and every Commander precon has a page of its own. A precon's page says how well its cards work together and which swaps make them work together more; a commander's shows what works with it, over every card a deck it leads could legally hold.

A page per commander	A page per precon

Start anywhere: Skullclamp · Krenko, Mob Boss · Atraxa, Praetors' Voice · Inalla, Archmage Ritualist · Edgar Markov · Yuna, Grand Summoner · every precon · or browse every commander.

Why you can trust it

It reads the cards, not the crowd. Nothing comes from card names, popularity or which cards show up together in other people's decks. Every claim comes from printed rules text, so a card no one plays and a staple in every deck are read the same way.

No AI when you analyse a deck. A model reads each card's oracle text exactly once, offline, and turns it into structured clauses. Everything after that is plain rules: the same deck always gets the same report, and every sentence traces back to a line on a card.

It says nothing rather than guess. A card it cannot read is listed as unread. A pairing it cannot verify is refused. A silent wrong answer is worse than a missing one.

And it is measured, not asserted. Every change to the engine is scored against a frozen panel of card pairs a person has judged by hand:

measure	result
synergy-claim precision	98.7% [97.1, 99.5] on 421 live claims
retention of pairs judged real	91.4% — 383 held, 36 lost
judged against	895 frozen card pairs, every claim hand-judged by a human
calibration corpus	71 real decks

Measured 2026-09-08 by npx tsx packages/instruments/src/panel-score.ts, which is free to re-run.

The two numbers only mean anything together. A tool that deleted every claim it was unsure of would score 100% on the first and collapse on the second. And retention is not recall: the panel was built from claims this engine already made, so it cannot see a pairing that was never claimed.

How it works

A pairing is found when something one card causes is something another card cares about: Krenko makes Goblin tokens enter, Impact Tremors cares about creatures entering, so the two pair. Farseek puts a land onto the battlefield, Enduring Courage cares about a creature entering, so they do not.

flowchart LR
  subgraph offline ["Once per card, offline"]
    A["Oracle<br/>text"] --> B["Segment"]
    B --> C["Normalize<br/>a model<br/>(paid)"]
    C --> D["Derive<br/>causes and<br/>cares about"]
  end
  subgraph browser ["In your browser"]
    E["Your<br/>deck"] --> F["Match"]
    F --> G["Report"]
  end
  D --> F

  style C fill:#7c2d12,color:#fff
  style D fill:#14532d,color:#fff
  style F fill:#1e3a8a,color:#fff

Only the red step costs money, and the project pays it once per card. The rest re-runs for free, which is why a wrong claim can be fixed and re-measured the same day.

The full story, for players and for engineers, is at edhseer.cards/how-it-works. The tour with the code is docs/HOW-IT-WORKS.md.

Honest limitations

• Corpus coverage. The corpus is 34,433 cards, of which 28,416 carry derived tags (measured 2026-09-09; 31,731 of the 31,829 commander-legal cards are read in full, the 98 left being persist-gate refusals). Cards outside that set form no pairings, and the report says so rather than quietly under-reporting. Their mana cost, type and text still count everywhere else.
• A pairing is found or not; it has no weight. Krenko handing Impact Tremors a token every turn and a lone creature entering once are the same pair. A supply/demand magnitude discount was built, swept across a 2-D parameter grid, and refused on measurement three separate times: it consistently penalised exactly the scarce payoffs it was meant to reward.
• Tutors and recursion pair only when they name a class. "Search your library for a Goblin" pairs with your Goblins and "return an artifact card" with your artifacts. "A card" or "a creature card" names most of a deck, so it pairs with nothing rather than with everything.
• The 71 calibration decks are one player's collection, not a metagame. Thresholds tuned against them are described as such wherever they appear.
• Mana simulation is a goldfish. No opponent, no interaction, no removal. Every figure it produces is a ceiling under a mana-maximising play policy, and says so.

For developers

Working, measured, and under active development: a deterministic engine with a regression panel, several ratchets, and a habit of recording the measurement beside every change.

package	what it does
@edh-seer/data	Scryfall + MTGJSON ingestion, MongoDB corpus, name resolution
@edh-seer/tagger	oracle text → clauses (the paid step) → derived tags (free)
@edh-seer/matcher	pairings, reasons, archetypes, mana math, build benchmarks, combos
@edh-seer/engine	scoring, ratings, per-card impact
@edh-seer/cli	terminal deck report
@edh-seer/instruments	the measuring tools: panel score, population diffs, recall samples
@edh-seer/web	the site: a React/Vite app that analyses in the browser, and the Pages Functions that prerender card pages
@edh-seer/import-worker	the Cloudflare Worker that imports Moxfield and Archidekt decks

Node 24 (.nvmrc; nvm use picks it up) is all the test suite needs:

npm install
npm test                  # 4,342 tests, green on 2026-09-25. No database, no network, no API key

Running the engine on real decks needs MongoDB holding the card corpus:

npx tsx packages/cli/src/main.ts <decklist.txt>                  # a deck report in the terminal
npx tsx packages/matcher/src/bin/build-static.ts                 # the card shards the site reads
npm run dev -w @edh-seer/web                                     # the site, on :5173

How to get the corpus, and everything else about running it, is in the runbook.

Documentation

Start at docs/, or go straight to what you need:

document	what it covers
How it works	the tour: a printed card to a synergy graph, with diagrams and a worked example
edhseer.cards/how-it-works	the same story on the site: a player's half, then an engineer's half, with the figures above held to this README by a test
Stage 1 — Segmentation	oracle text into numbered clauses, free
Stage 2 — Normalization	the model, once per card. The only paid step
Stage 3 — Derivation	clauses into game events, plus what no card prints
Stage 4 — Matching	what one card causes against what another cares about, and every reason a claim is refused
Schema reference	every vocabulary, constant and field. Generated from the source and gated by a test
Runbook	what to run, what it costs, how to measure it afterwards
Engineering log	what was measured, on what day, and what it cost

Contributing: CONTRIBUTING.md. Found a wrong pairing? Report it with the two card names and the sentence the report printed. It is the most useful issue this project can receive.

License

MIT — see LICENSE.

Card data from Scryfall and MTGJSON. Combo data from Commander Spellbook. Magic: The Gathering is a trademark of Wizards of the Coast. This project is unaffiliated with them.