kasper-cbx/mtg-search-index

MTG Search Index

The card corpus and search index of mtg-search, an MCP server that gives an AI agent semantic search over every Magic: The Gathering card for deck building. Ask for "counter target spell unless its controller pays, legal in pauper" and it returns the cards whose rules text does that, filtered by colour identity, format, type and mana value.

• One entry per distinct card (Scryfall oracle_id), whatever its printings: 34,646 cards, 4,579 Tagger function tags and 78,973 rulings.
• Official precon decklists: 1,871 preconstructed decks from MTGJSON (MIT), 197 of them Commander decks, keyed by oracle_id. The server's get_deck tool loads them by deck or commander name.
• Snapshot: Scryfall oracle_cards updated_at 2026-10-15T21:01:57.844+00:00, MTGJSON 5.3.0+20261015, tagged scryfall-20261015210157.
• Search: lightonai/mLateOn late interaction (multi-vector, MaxSim), exact over all cards: no approximate index. Queries are encoded with mLateOn's published fp32 ONNX model at the revision the cards were encoded with. The shipped server returns exactly the same top 10 as the benchmark reference on all 363 benchmark queries.
• Evaluation data that never reaches the server: eval/popularity.parquet (EDHREC rank), used only to analyse popularity bias.

Related datasets (same project; both were built on the card snapshot scryfall-20261014090155):

• kasper-cbx/mtg-search-bench: the search benchmark (363 queries), the 113 agent search tasks, and graded relevance labels.
• kasper-cbx/mtg-precon-bench: the precon deck benchmark (player add/cut rates from EDHREC, and deck-fit grades).

This dataset contains card names and rules text derived from Scryfall bulk data (© Wizards of the Coast) and decklists from MTGJSON (MIT), published as the search index of the mtg-search MCP server. See Data sources and terms.

Citation

Cite a tagged snapshot, so the cards and embeddings are fixed:

@misc{mtgsearch_index_2026,
  author       = {Gr{\o}ntved, Kasper},
  title        = {{MTG Search Index}: Scryfall Oracle cards with mLateOn late-interaction embeddings},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/kasper-cbx/mtg-search-index}},
  note         = {Revision scryfall-20261015210157 (Scryfall oracle\_cards 2026-10-15). Card data from Scryfall, decklists from MTGJSON, \textcopyright{} Wizards of the Coast}
}

---

Data sources and terms

Scryfall bulk data

All card data comes from Scryfall bulk data. It is fetched through https://api.scryfall.com/bulk-data/{type} (with a User-Agent: mtg-search/0.1 header), then each jsonl_download_uri (gzipped JSONL) is streamed. The three files used, as recorded in data/manifest.json:

Bulk type	Scryfall updated_at	File
oracle_cards	2026-10-15T21:01:57.844+00:00	https://data.scryfall.io/oracle-cards/oracle-cards-20261015210157.jsonl.gz
oracle_tags	2026-10-15T21:00:33.145+00:00	https://data.scryfall.io/oracle-tags/oracle-tags-20261015210033.jsonl.gz
rulings	2026-10-15T21:00:37.150+00:00	https://data.scryfall.io/rulings/rulings-20261015210037.jsonl.gz

• Oracle tags are community annotations from Scryfall Tagger, such as sacrifice-outlet, ramp and removal-creature.
• Rulings are the official Wizards of the Coast rulings plus Scryfall's notes, as Scryfall publishes them.

MTGJSON decklists

data/precons.parquet is built from MTGJSON's AllDeckFiles.tar.gz (build 5.3.0+20261015), downloaded with an identifying User-Agent and checked against its published .sha256; each refresh requests it again only when MTGJSON has a newer build. Only deck names, types, set codes, release dates, the publisher's decklist link, and per card the oracle_id, name, count and section are kept: MTGJSON's prices, purchase links, ranks and printing fields are dropped. Card bundles that are not decks (Secret Lair drops, land packs, MTGO redemption lists, Deck Builder's Toolkits and similar) are left out.

MTGJSON's data is MIT licensed:

MIT License

Copyright © 2018-Present Zachary Halpern

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

The card names in the decklists are © Wizards of the Coast, as below.

Terms, copyright and affiliation

• Card content. Card names, rules text, mana costs and type lines are © Wizards of the Coast LLC. They are used under the Wizards of the Coast Fan Content Policy. This is unofficial Fan Content, not approved or endorsed by Wizards. Portions of the materials used are property of Wizards of the Coast. © Wizards of the Coast LLC.
• Scryfall. Use of Scryfall data follows the Scryfall API terms. This dataset is the search index of the mtg-search tool: the card tables ship together with the embeddings and tags they were built for. The license: other in the metadata reflects that it is derived data, not openly licensed.
• No affiliation. This project is not affiliated with, endorsed by, or sponsored by Scryfall, MTGJSON or Wizards of the Coast.
• Images. Card images are not stored here. image_uri and scryfall_uri are links to Scryfall.

Model

Model	Role	License	Revision
lightonai/mLateOn	Document embeddings and query encoder (fp32 model.onnx)	Apache-2.0	bbd883ae1fadb563056ff73b1061ae8172a3fcda

---

Repository layout

Path	Size	Description
data/cards.parquet	8.5 MB	One row per card: text, costs, types, legalities, tags, deck-building fields. Config cards.
data/tags.parquet	3.6 MB	One row per Tagger function tag, with its direct taggings. Config tags.
data/rulings.parquet	2.3 MB	One row per ruling. Config rulings.
data/precons.parquet	0.9 MB	One row per official preconstructed deck, from MTGJSON. Config precons.
data/manifest.json	1 KB	Snapshot metadata: Scryfall updated_at and URIs per bulk file, row counts, build time, and the content fingerprint.
embeddings/late/lightonai__mLateOn/tokens.npy	557.9 MB	float16 (2,179,160, 128): per-token card embeddings, concatenated in card order.
embeddings/late/lightonai__mLateOn/offsets.npy	0.3 MB	int64 (34,647,): card i is tokens[offsets[i]:offsets[i+1]].
embeddings/late/lightonai__mLateOn/meta.json	1 KB	Model, revision, prompts, alignment hash, build info.
eval/popularity.parquet	0.8 MB	oracle_id, edhrec_rank. Evaluation only: the server never loads, shows or ranks by it, and the Docker image does not contain it. Config popularity.

---

Schema

data/cards.parquet (34,646 rows, 46 columns, sorted by name)

Column	Type	Meaning
oracle_id	string	Scryfall oracle id. The primary key, and the join key for every other file.
name	string	Card name. Multi-face cards use Front // Back.
face_names	list[string]	Name of each face; a single-faced card has one.
doc_text	string	The text that is embedded and indexed (see below).
mana_cost	string	e.g. {1}{U}{U}. Faces are joined with // when there is no top-level cost.
cmc	float64	Mana value.
pips_W … pips_G, pips_C	int16	Coloured-symbol counts with devotion semantics: hybrid {G/W} counts for both, and {2/W} and Phyrexian {W/P} count for W. The value is the maximum over faces.
generic	int32	Sum of the numeric generic symbols (maximum over faces).
has_x / has_hybrid / has_phyrexian	bool	Any face has {X}/{Y}/{Z}, a hybrid symbol, or a Phyrexian symbol.
face_costs	list[struct]	The same cost breakdown per face (name, mana_cost, pips, generic, flags).
colors	list[string]	Card colours (the union over faces if needed), in WUBRG order.
color_identity	list[string]	Commander colour identity.
type_line	string	Full type line; faces joined with //.
supertypes / types / subtypes	list[string]	Parsed from type_line, across all faces.
keywords	list[string]	Scryfall keyword abilities and actions.
legalities	struct of 23 strings	Per format: legal / not_legal / restricted / banned. The formats are alchemy, brawl, commander, competitivebrawl, duel, future, gladiator, historic, legacy, modern, oathbreaker, oldschool, pauper, paupercommander, penny, pioneer, predh, premodern, standard, standardbrawl, timeless, tlr and vintage.
game_changer	bool	Card is on the Commander "Game Changers" list.
is_funny	bool	Un-set / playtest card that is not tournament legal (see filtering).
layout	string	Scryfall layout (normal, transform, modal_dfc, split, adventure, saga, …).
set / set_type / released_at	string	Set code, set type and release date of the representative printing.
scryfall_uri	string	Card page on Scryfall.
image_uri	string	Scryfall normal image URL (front face for multi-face cards).
tags	list[string]	Slugs of the Tagger oracle tags applied directly to this card (no ancestor closure).
rarity	string	Rarity of the representative printing.
produced_mana	list[string]	Mana the card can produce, card-level (e.g. ["C"] for Sol Ring).
is_permanent	bool	Any face is an artifact, creature, enchantment, land, planeswalker or battle.
is_legendary	bool	Legendary supertype.
can_be_commander	bool	Legendary creature on the front face, or text saying it "can be your commander".
partner_kind	string (nullable)	How it pairs as a second commander: partner, partner with, friends forever, partner—<group>, choose a background, background, doctor's companion, doctor.
power / toughness / loyalty / defense	string (nullable)	Front-face stats. These are strings because values like *, 1+* and X occur.
faces	list[struct]	Per face: name, mana_cost, type_line, oracle_text, power, toughness, loyalty, defense, is_permanent.
reserved	bool	On the Reserved List.

Prices are not stored (removed in manifest schema_version 2): they change daily, the server does not use them, and they would make every snapshot look new.

data/tags.parquet (4,579 rows)

Column	Type	Meaning
slug	string	Tag slug, as in Scryfall's otag:<slug>.
label	string	Display name.
description	string	Tagger description (often empty).
aliases	list[string]	Alternative slugs.
parent_slugs / child_slugs	list[string]	Tag hierarchy. Relevance in the eval uses the closure over child_slugs.
card_count	int64	Number of cards in this corpus tagged directly. 4,410 tags have at least one.
oracle_ids	list[string]	Directly tagged oracle ids, restricted to this corpus.

data/rulings.parquet (78,973 rows)

Column	Type	Meaning
oracle_id	string	Card (join to cards).
source	string	wotc or scryfall.
published_at	date	Ruling date.
comment	string	Ruling text.

Only rulings for cards that appear in cards.parquet are kept. They are sorted by (oracle_id, published_at).

data/precons.parquet (1,871 rows)

Every preconstructed deck in MTGJSON's AllDeckFiles except card bundles: Jumpstart (570), Theme Deck (220), Commander Deck (197), Intro Pack (167), Arena Starter Deck (101) and 25 more types, released 1993-12-10 to 2026-10-02. Sorted by release_date (newest first), then name.

Column	Type	Meaning
file	string	MTGJSON deck file name, e.g. CallingAllAngels_FDC (mtgjson.com/api/v5/decks/<file>.json).
name	string	Deck name.
type	string	MTGJSON deck type: Commander Deck, Jumpstart, Duel Deck, Historic Brawl Precon Deck, …
set_code	string	Set code of the product.
release_date	string	Release date (YYYY-MM-DD).
source	string (nullable)	The publisher's decklist page, when MTGJSON has one.
commanders	list[string]	Commander names (empty for non-Commander decks).
cards	list[struct]	Per card: oracle_id (join to cards), name, qty, section (commander, main or sideboard). Foil and non-foil copies of a card are merged.

50,663 card rows; every one has an oracle_id. 164 of them (Planechase planes and Theros Challenge Deck cards in box sets) are cards that cards.parquet leaves out by layout.

eval/popularity.parquet (34,646 rows, evaluation only)

Column	Type	Meaning
oracle_id	string	Card (join to cards).
edhrec_rank	int64 (nullable)	EDHREC's overall popularity rank at snapshot time (1 = most played; null for 2,305 cards). Used only to measure popularity bias in evaluations. It is not part of the content fingerprint, so a rank change never triggers a refresh.

Content fingerprint

data/manifest.json carries deterministic hashes of the tables (fingerprint_version 1, computed by mtg_search.ingest.fingerprint). Each hash is SHA-256 over the selected columns of every row as canonical JSON (sorted keys), one line per row, in cards.parquet order. Every cards.parquet column belongs to exactly one group.

Field	Covers	A change means
text_hash	oracle_id, doc_text	The embeddings must be re-encoded.
oracle_ids_sha256	oracle_id order (same digest as the embedding meta.json files)	Cards were added, removed or reordered: re-encode.
gameplay_hash	oracle_id, name, face_names, faces, mana_cost, face_costs, pips_W … pips_C, generic, has_x, has_hybrid, has_phyrexian, cmc, colors, color_identity, type_line, supertypes, types, subtypes, keywords, legalities, game_changer, reserved, layout, produced_mana, is_funny, is_permanent, is_legendary, can_be_commander, partner_kind, power, toughness, loyalty, defense, tags	Rules facts changed (e.g. a ban or new tags): publish new tables, reuse the embeddings.
aux_hash	all of tags.parquet and rulings.parquet (row order ignored)	Publish new tables, reuse the embeddings.
precons_hash	all of precons.parquet, as canonical JSON per row in file order (mtg_search.precons.fingerprint)	An official decklist changed or a deck was added: publish new tables, reuse the embeddings.
(not hashed)	image_uri, scryfall_uri, and the representative printing's set, set_type, released_at, rarity	Nothing: these are volatile and never trigger a refresh on their own.

Current values: text_hash 18caede2…, gameplay_hash e4399316…, aux_hash df8d47a7…, precons_hash df3c39a0…, oracle_ids_sha256 1bff7b49…. The manifest's precons field records the MTGJSON build and the deck count.

How doc_text is built

src/mtg_search/doctext.py renders one block per face:

{name} {mana_cost}
{type_line}
{oracle_text}
{power/toughness | Loyalty: n | Defense: n}

• Multi-face cards (transform, MDFC, split, adventure, flip, …) repeat the block for each face, separated by a line containing only //.
• Keywords: the card-level keywords are appended once, as Keywords: Flying, Transform.
• Left out: empty lines and flavor text (never included). Tags are not part of doc_text either. Example:

Delver of Secrets {U}
Creature — Human Wizard
At the beginning of your upkeep, look at the top card of your library. ...
1/1
//
Insectile Aberration
Creature — Human Insect
Flying
3/2
Keywords: Flying, Transform

Filtering rules

Rules applied by src/mtg_search/ingest.py to the 38k-object oracle_cards file:

• Excluded layouts: token, double_faced_token, emblem, art_series, planar, scheme, vanguard, front_card.
• Excluded set type: memorabilia.
• Kept but flagged: Un-cards. is_funny = True only when set_type == "funny" and the card is not legal or restricted in vintage. Eternal-legal cards from funny sets, such as Unfinity's non-acorn cards, therefore stay False. This gives 1,420 cards flagged, out of 1,467 with set_type == "funny". The MCP server's tools always exclude them.
• Join scope: tags and rulings are restricted to the oracle ids that remain after filtering.

---

Embedding artifacts

The embeddings were computed by jobs/encode.py, a PEP-723 uv script, on Hugging Face Jobs (t4-small, Tesla T4), with torch 2.14.1+cu130, transformers 5.20.0 and sentence-transformers 6.1.0. The input is data/cards.parquet (oracle_id, doc_text).

Artifact	Format	Dim	Query prompt	Doc prompt	Doc max len	Doc tokens p50 / max	Truncated	Encode time (T4)
late lightonai/mLateOn	tokens.npy fp16 + offsets.npy int64; 2,179,160 tokens	128	[Q]	[D]	8192	59 / 398	0	213 s

Row alignment. Row i of the embeddings belongs to row i of data/cards.parquet (sorted by name). meta.json records oracle_ids_sha256, the SHA-256 of all oracle_ids in that order, each followed by \n. Check it before using the embeddings with a cards.parquet from another revision:

import hashlib, json, numpy as np, polars as pl
meta = json.load(open("embeddings/late/lightonai__mLateOn/meta.json"))
ids = pl.read_parquet("data/cards.parquet", columns=["oracle_id"])["oracle_id"]
h = hashlib.sha256(); [h.update(i.encode() + b"\n") for i in ids]
assert h.hexdigest() == meta["oracle_ids_sha256"]  # 1bff7b49... for this snapshot

tokens  = np.load("embeddings/late/lightonai__mLateOn/tokens.npy", mmap_mode="r")
offsets = np.load("embeddings/late/lightonai__mLateOn/offsets.npy")
card_i  = tokens[offsets[i]:offsets[i + 1]]          # (n_tokens_i, 128) float16

Search. A query is encoded with the [Q] prompt by mLateOn's fp32 model.onnx (revision above), and each card is scored by MaxSim: for every query token, the best-matching card token, summed. All 34,646 cards are scored exactly, as one matrix product; at this size an approximate index buys nothing. Filters restrict the cards that compete before ranking. Measured parity with the benchmark reference on all 363 queries: 100% top-1 and top-10 agreement, identical nDCG@10.

---

Evaluation summary

The search benchmark (kasper-cbx/mtg-search-bench, config queries) has four query sets, scored against all 34,646 cards with funny cards excluded and per-query filters applied: tags (65, from Tagger labels), handwritten (46, rules-text wording, regex labels), paraphrase (222, the same needs in players' words), deckneeds (30, open-ended deck needs). nDCG@10 × 100:

System	tags	handwritten	paraphrase	deckneeds	para+deck
mLateOn (this index)	60.5	70.7	11.9	21.1	13.0
dense (embeddinggemma-2)	42.8	36.7	9.1	17.9	10.1
sparse (opensearch doc-v3-gte)	40.0	47.3	4.9	19.6	6.7
BM25	36.4	61.1	2.5	8.5	3.2

The other systems' embeddings are in revisions up to 9776f07 (see History). The answer lists come from tags and patterns, so absolute scores are lower bounds; compare systems with each other. With player-worded requests rewritten as rules text, mLateOn reaches 75.5. Agent-level results (search vs grep vs reading every card, deck building) are in the project's findings and experiment registry.

---

Update process

scripts/refresh.py keeps this repo in sync with Scryfall, run by hand when needed (e.g. after a set release or a ban announcement) through a GitHub Action, .github/workflows/refresh.yml. Prices and popularity never trigger an update; legality and text changes always do.

Tier	When	What happens
No-op	All content hashes equal: only volatile fields changed.	Nothing is uploaded or tagged.
Metadata-only	gameplay_hash, aux_hash or precons_hash changed, but text_hash and oracle_ids_sha256 did not (bans, tags, rulings, precon lists, …).	Upload data/*, check that the embeddings still match oracle_ids_sha256, tag. No re-encoding.
Re-encode	text_hash or oracle_ids_sha256 changed (errata, new cards).	Upload data/*, re-encode mLateOn on an HF Job, verify alignment, run the eval smoke test (fails if mLateOn nDCG@10 drops more than 3 points), tag.

Every publishing refresh tags the new snapshot scryfall-<YYYYMMDDHHMMSS> (from Scryfall's updated_at), and tags the snapshot it replaces if it has no tag yet. To reproduce or cite a build, pin revision="scryfall-…".

---

Usage

The repo is public, so no login is needed.

from datasets import load_dataset
cards = load_dataset("kasper-cbx/mtg-search-index", "cards", split="train", revision="scryfall-20261015210157")

import polars as pl
from huggingface_hub import hf_hub_download

path = hf_hub_download("kasper-cbx/mtg-search-index", "data/cards.parquet", repo_type="dataset",
                       revision="scryfall-20261015210157")
cards = pl.read_parquet(path)
cards.filter(
    pl.col("legalities").struct.field("pauper").is_in(["legal", "restricted"]),
    pl.col("color_identity").list.eval(pl.element().is_in(["U", "B"])).list.all(),
    ~pl.col("is_funny"),
).select("name", "mana_cost", "type_line")

The MCP server (Docker; the build pulls this dataset with HF_TOKEN as a BuildKit secret):

git clone https://github.com/kasperg3/mtg-search && cd mtg-search
docker/build.sh --revision scryfall-20261015210157
claude mcp add mtg-search -- docker run -i --rm mtg-search

The server's read-only tools: search_cards (1–4 queries, filters for colour identity, format, types and mana value), get_card, find_similar and, from mtg-search-mcp 0.2.0, get_deck (a Moxfield or Archidekt link, an official precon from data/precons.parquet, or a pasted list, checked against its format's deck rules). Results are ordered by text relevance only. Tool output carries rules and print facts only: no popularity, price or other crowd-derived data. game_changer is shown because it is the official Commander bracket designation.

---

Limitations

• Tags are incomplete and noisy. Tagger is community-maintained: 204 cards have no tags, newer or obscure cards are often under-tagged, and some tags are loose. tags holds direct taggings only; take the closure over child_slugs yourself.
• One printing per card. rarity, set, released_at and image_uri come from the representative printing Scryfall's oracle_cards file picks, not the first printing. No prices.
• English Oracle text only. No flavor text, art, artist or printing-specific text.
• Un-cards are present, flagged with is_funny.
• Snapshot drift. Card text and legality change with bans and errata: pin a scryfall-… tag, and check oracle_ids_sha256 before mixing embeddings and tables from different revisions.
• Derived data. Card text is © Wizards of the Coast; see Data sources and terms.

History

• Up to revision 9776f07: this repo also held the retrievers compared in the search benchmark and then retired: dense google/embeddinggemma-2, learned sparse opensearch-neural-sparse-encoding-doc-v3-gte and Linkup-Platform/linkup-sparseup-embed-v1, late answerdotai/answerai-colbert-small-v1, and a re-quantized int8 mLateOn query encoder for the earlier NextPlaid image. They were removed from main when the server moved to exact mLateOn search. Pin 9776f07 to reproduce those experiments.
• scryfall-20261015210157 (2026-10-16): the first snapshot with data/precons.parquet (MTGJSON's official decklists) and precons_hash in the manifest. The search and precon benchmarks stay on scryfall-20261014090155.
• Schema 2 (manifest schema_version 2): prices were dropped. edhrec_rank moved from data/cards.parquet to eval/popularity.parquet, so the server's tables carry no popularity data.