Lorcana MCP Server

An MCP server for searching and aggregating Disney Lorcana cards.

Startup behavior

On startup, the server fetches a JSON list of cards from https://danielenricocahall.github.io/lorcana-mcp/allCards.json. The snapshot is refreshed daily by data_pipeline/fetch_cards.py, which pulls from the Lorcast API, normalizes each card into our internal schema, and publishes the list to the gh-pages branch. The middle layer insulates running containers from Lorcast's availability and rate limits — the runtime never calls Lorcast directly.

Cards are kept in-memory as a Python list for fast filtering. With ~2,270 unique cards (each carrying a printings array for its alternate sets/rarities) this is lightweight and requires no external database. A local JSON file cache (LORCANA_CACHE_PATH, default cards.json) lets the server skip the network fetch on subsequent startups.

Startup data loading is controlled by:

• LORCANA_REFRESH_ON_STARTUP:
  • true: always fetch from API and repopulate storage
  • false: use existing cache if available
• LORCANA_SKIP_IF_DB_EXISTS:
  • true (default): skip API fetch if the cache file already contains cards
  • false: fetch and repopulate

Quick start (no clone required)

The server is published to GHCR and the MCP Registry. Pull and run it directly:

docker pull ghcr.io/danielenricocahall/lorcana-mcp:latest

docker run --rm -i ghcr.io/danielenricocahall/lorcana-mcp:latest

To persist the card cache across container restarts, mount a volume:

docker run --rm -i \
  -e LORCANA_CACHE_PATH=/data/cards.json \
  -e LORCANA_SKIP_IF_DB_EXISTS=true \
  -v lorcana_mcp_data:/data \
  ghcr.io/danielenricocahall/lorcana-mcp:latest

Run locally (stdio MCP)

uv run python main.py

Docker

Build image

docker build -t lorcana-mcp:latest .

Run as stdio MCP server

docker run --rm -i lorcana-mcp:latest

Docker Compose

Start with compose

docker compose build
docker compose run --rm -T lorcana-mcp

Notes:

• No port is exposed; MCP communication is over stdio.
• Use a volume to persist the JSON cache across restarts.

Config

• LORCANA_API (default: https://danielenricocahall.github.io/lorcana-mcp/allCards.json)
• LORCANA_CACHE_PATH (default: cards.json) — local file for caching fetched cards
• LORCANA_HTTP_TIMEOUT_SECONDS (default: 60)
• LORCANA_REFRESH_ON_STARTUP (false default)
• LORCANA_SKIP_IF_DB_EXISTS (true default)

MCP client setup examples

Local process (Claude Desktop-style)

{
  "mcpServers": {
    "lorcana": {
      "command": "uv",
      "args": ["run", "python", "/absolute/path/to/lorcana-mcp/main.py"]
    }
  }
}

Published image — GHCR (Claude Desktop-style, no clone required)

{
  "mcpServers": {
    "lorcana": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "ghcr.io/danielenricocahall/lorcana-mcp:latest"
      ]
    }
  }
}

Docker process (Claude Desktop-style, locally built)

{
  "mcpServers": {
    "lorcana": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "lorcana-mcp:latest"
      ]
    }
  }
}

Docker Compose process (Claude Desktop-style)

{
  "mcpServers": {
    "lorcana": {
      "command": "docker",
      "args": ["compose", "run", "--rm", "-T", "lorcana-mcp"]
    }
  }
}

Via the Claude CLI — published image (global, no clone required)

claude mcp add --scope user \
  -- lorcana docker run --rm -i \
  ghcr.io/danielenricocahall/lorcana-mcp:latest

Via the Claude CLI — locally built

claude mcp add --scope user \
  -- lorcana docker run --rm -i lorcana-mcp:latest

Example questions

Once connected to an MCP client, you can ask natural language questions like:

Card lookup

• "Show me all cards named Moana"
• "What does the card Maui - Hero to All do?"
• "Find all legendary amber cards"

Deck building

• "What are the cheapest ruby characters with at least 3 strength?"
• "Show me inkable sapphire cards that cost 4 or less"
• "Find steel characters with 5 or more willpower"
• "What 3-lore characters exist in emerald?"

Keyword & ability search

• "How many Singer cards cost exactly 5?"
• "How many Evasive characters are there in the first set?"
• "How many ruby cards have Reckless?"
• "Find all cards with Ward in their text"
• "Show me Shift cards in amethyst"

Stats & aggregations

• "How many cards are in each set?"
• "What's the color distribution across all cards?"
• "What are the most common traits?"
• "Show me the ink curve — how many cards exist at each cost?"
• "How many legendary cards are inkable?"

Cross-filter queries

• "How many amber characters have 3 or more lore?"
• "Find cheap (cost 2-3) characters with high strength (4+) in steel"
• "How many cards in set 1 have Evasive and cost less than 4?"

Note: For plain keyword queries (Evasive, Bodyguard, Shift, etc.) use the keyword parameter — it filters against the structured ability list and is more reliable than substring search. For value-specific queries like Singer 5 or Resist +2, use body_text (keyword values live in the card's full text, not the ability list).

MCP tools

• search_cards — filter and retrieve card objects (supports response_format="toon" for ~10% fewer tokens)
• count_cards — count cards matching a filter without returning full objects
• aggregate_cards — card counts grouped by cost (ink curve), rarity, color, set_code, or type
• resolve_card — fuzzy-match an informal/partial/misspelled card name to the closest cards (returns full card data)
• top_traits — most common traits across all cards
• export_deck — render a deck as a Dreamborn/Pixelborn-compatible text deck list
• import_deck — parse a Dreamborn/Pixelborn-style deck list, returning resolved cards plus any unresolved lines with fuzzy candidates
• validate_deck — check a deck against the format rules (≥60 cards, max 4 copies, ≤2 inks); returns {legal, total_cards, inks, violations}
• deck_stats — compute ink curve, color split, inkable count, and type breakdown for a deck
• server_status — startup metadata (card count, config)

MCP prompts

• build_deck(colors, playstyle="balanced") — guides the model through assembling a legal Lorcana deck (60-card minimum, ≤2 inks, max 4 copies of any card) for the requested color(s) and playstyle (aggressive / control / lore-race / balanced). Uses the search/aggregate tools above plus the rules embedded in the server instructions.

TOON response format

search_cards accepts a response_format argument:

• "json" (default) — list of card objects, unchanged from prior versions.
• "toon" — a TOON string with one column header line and one row per card, encoded by the toons Rust-backed library (the official community reference implementation).

Example (search_cards(name="elsa", limit=2, response_format="toon")):

cards[2]:
  - id: crd_01c4835a62df4960bb973aeff81f2bb2
    name: Elsa
    version: Ice Maker
    full_name: Elsa - Ice Maker
    cost: 7
    ...
    printings[3]{set_code,set_name,number,rarity}:
      "7",Archazia's Island,69,Super Rare
      C2,Lorcana Challenge Year 3,2,Promo
      C2,Lorcana Challenge Year 3,6,Promo
  - id: crd_04bca46a8e2d4e9ba0fbdbfc6c99e51e
    name: Elsa
    ...

The outer cards[2]: falls back to YAML-style per-card blocks (rather than a single tabular table) because card shapes vary — Actions and Items don't carry strength/willpower/lore, for example. The inner printings[N]{...}: block is fully tabular since every printing has the same four fields.

Benchmark

Measured with benchmarks/bench_toon.py against the live ~2,270-card dataset (post-consolidation), tokenizing with tiktoken cl100k_base (used as a proxy for Claude's tokenizer):

query	rows	JSON tokens	TOON tokens	Δ
color="amber", limit=200	200	43,672	39,282	−10.1%
color="ruby", limit=50	50	10,446	9,464	−9.4%
card_type="action", limit=50 (sparse cols)	50	10,150	9,265	−8.7%
body_text="when", limit=50 (long full_text)	50	11,574	10,380	−10.3%
name="elsa", limit=20	14	3,456	2,925	−15.4%
total		79,298	71,316	−10.1%

Note: TOON's relative savings are smaller here than they were before the printings consolidation (pre-PR-#29 the same queries showed ~50% reductions). That gap is structural to the nested printings array — TOON's columnar encoding wins on the top-level fields but falls back to JSON-style encoding inside the per-printing entries, so the array dilutes the relative gain. Absolute token counts are still down meaningfully versus the equivalent count of pre-consolidation rows since each unique card is now represented once with a small printings list rather than as 1-3 separate full rows.

Reproduce with PYTHONPATH=. uv run python benchmarks/bench_toon.py (requires a populated cards.json cache).

Disclaimer

This is a personal, unofficial fan and engineering project. It is not affiliated with, endorsed by, sponsored by, or reviewed by Disney, Ravensburger, or the Disney Lorcana TCG team. I worked only with publicly available/community data sources. All Disney Lorcana TCG names, card text, trademarks, and related intellectual property belong to Disney and Ravensburger. This project is non-commercial and reflects my personal views only, not those of my employer.