NeoSatoshi Card Recognition MCP server
MCP server for the NeoSatoshi card recognition API: lets Claude, Cursor, Codex and other MCP clients identify Pokémon, One Piece, Disney Lorcana, Yu-Gi-Oh!, Magic: The Gathering, Star Wars: Unlimited, Digimon Card Game and football (soccer) cards in a photo (binder pages work, football one card per photo) and get set, number, rarity, Cardmarket ID and daily prices as JSON (football: player, release and parallel), and search the card catalog.
- Transport: stdio
- npm package:
neosatoshi-card-mcp, binary neosatoshi-card-mcp
- Node.js 18 or newer
Tools
recognize_cards
| include_images | boolean | Also return the catalog image of up to 10 recognised cards (from our CDN, at most 4 MB together), so the assistant can compare them with the photo. Default false. |
Give exactly one of image_path, image_url, image_base64. The result has a short text summary per card, the full API response as JSON text, and the same response as structuredContent. API errors come back as tool errors with the API's error code and message.
Example summary line:
Digimon line (alternative arts share the card number; the printing id and its variant follow the set):
Digimon Card Game (digimon), English and Japanese: recognised by card number and name. When the picture cannot tell the
alternative arts apart, print_unconfirmed is true, card_id is dg-c-<number> without a price, and the printings
come as alternatives, each with its price, to pick with correct_card. Cardmarket prices (EUR) for English and Japanese
printings, TCGplayer (USD) for English only. Older Digimon products (the 1999 card game, the 2000 trading cards) are not
recognised.
Football line (no prices yet, the parallel is a suggestion a person should confirm):
When the client sends a progress token (Claude Code, Claude Desktop and most clients do), the detection is streamed and each card is reported as soon as it is recognised: on an 8 card Lorcana page the first card came after about 8 seconds, all eight after 11.
The API identifies cards; it does not grade condition (a PSA, BGS or CGC slab, for football also SGC, comes back with the grade read from the label). One card usually takes about 4 seconds, about 10 to 12 seconds on the first request after a quiet period; a binder page with 9 cards about 17 seconds; a football card about 6 to 25 seconds. The client waits up to 120 seconds.
recognize_folder
One detection per photo, one credit per detected card. Existing files are never overwritten. It stops early on errors that would fail every photo (no credits, invalid key, demo limit). CSV columns: file, card_index, name, tcg, language, set_name, set_code, number, rarity, cardmarket_id, cardmarket_trend_eur, cardmarket_avg30_eur, tcgplayer_market_usd, confidence, graded, season, subset, parallel, request_id. Cells that start with =, +, - or @ get a leading ' so a spreadsheet does not run them.
search_cards
Search the card catalog; results are full card objects like a detection returns (set, number, rarity, Cardmarket
product ID, daily prices, images), at most 10, with has_more but no total. With request_id and card_index from
recognize_cards it is a free correction search (5 per detected card within 24 hours, at most 5 results). Without
them it is the paid card search: 0.5 credits when it returns a card.
Football: {"tcg": "football", "q": "kane chrome 2023"}, only q (the player, optionally with set or season words), no
language, set or number, no prices yet. Each result carries ref for get_card.
correct_card
Correct a detected card to the right catalog card (card_id from a correction search or from the card's
alternatives). Free, within 24 hours, 3 times per card, only for the key that made the detection.
get_card
One card by its card_id, for example to refresh a price. 0.5 credits when the card exists. For a wrong detection
use search_cards with request_id and correct_card instead, both free.
Football: pass ref from the search_cards result (same key, within 24 hours). The card then comes with all parallels
of its subset and their print runs. Without a valid ref the call is refused and free.
export_cards
The cards of up to 20 detections (request_ids of recognize_cards or recognize_folder, within 48 hours) as the
same file the NeoSatoshi website exports, built by the website's own export code: your condition for new cards,
corrections made with correct_card, your saved presets, price rules and templates apply.
format csv/json (default csv): the custom export. preset (a saved preset id), or fields and json_mode
(api = the documented card shape, custom = your fields).format tcgpowertools: the TCGPowerTools CSV for Cardmarket (your TCGPowerTools rules and saved comment; comment).format shopify: the Shopify product CSV, prices in the store currency (store_currency, else read from your
connected store; variant_mode).format ebay: the eBay Seller Hub CSV for one site (marketplace, e.g. EBAY_DE; one_row_per_card).- For the last three:
price_source, min_price, markup_percent and rounding, like the buttons of the website's
pricing step (a markup or rounding replaces your saved price rules; markup_percent 0 = none). save_to_folder writes a new file into an allowed folder (never overwrites), otherwise
the file comes back in the answer (up to 100 KB). Free, own key only (demo detections keep no cards), not for football
yet. The answer names cards it left out (skipped) and cards whose price depends on a special version (warnings).
list_export_presets
Your saved export presets (plus the built-in recommended CSV) and every field id export_cards takes. Free.
list_supported_games
Games, card languages, tcg values, limits and card search prices. No API call, no cost.
Cost
Configuration
Setup
Claude Code
With a directory restriction:
Leave out -e NEOSATOSHI_API_KEY=... to try it with the demo key.
Claude Desktop, one click
Download neosatoshi-card-mcp.mcpb and open it: Claude Desktop installs the extension and asks for the optional API key and an optional allowed folder. Node.js comes with Claude Desktop.
Build it yourself with npm run build:mcpb (needs the dev dependencies).
Claude Desktop, by hand
Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json, Windows: %APPDATA%\Claude\claude_desktop_config.json):
Cursor
~/.cursor/mcp.json (global) or .cursor/mcp.json (project):
Codex CLI
~/.codex/config.toml:
From source
Point any client at node /absolute/path/to/mcp-server/dist/index.js instead of npx.
Security
- Files: only regular files are read. Directories, devices, sockets and pipes are refused. Symlinks are resolved first and the real target is checked. Only
.jpg, .jpeg and .png whose content starts with JPEG or PNG magic bytes are accepted, at most 12 MB (football 4 MB).
NEOSATOSHI_ALLOWED_DIRS is checked against the resolved real path, so ../ and symlinks cannot escape it. Entries that do not exist grant nothing.- Network: the server only calls
https://europe-west1-neosatoshi.cloudfunctions.net/detectApi and .../cardsApi, and with include_images our card image CDN https://d1m1gwcombja0u.cloudfront.net (other image hosts are skipped).
- Writing: only
recognize_folder writes, and only new files named neosatoshi_cards_<date>-<time>.csv / .json into the folder it read (same NEOSATOSHI_ALLOWED_DIRS check). image_url is passed to the API, which fetches and validates it; the server itself never fetches URLs.
- No shell execution.
- The API key is only sent in the
Authorization header. It is never logged or returned, and it is scrubbed from error messages.
Development
Publishing notes
server.json is for the official MCP registry (https://registry.modelcontextprotocol.io). Its name matches the mcpName in package.json (io.github.Cookiebreed/card-recognition-mcp); the registry checks GitHub ownership for io.github.* names.smithery.yaml is for Smithery.mcpb/manifest.json and scripts/build-mcpb.mjs build the Claude Desktop extension; upload neosatoshi-card-mcp.mcpb to each GitHub release under that exact name.
License
MIT for this code (see LICENSE). The API is a commercial service governed by https://neosatoshi.com/termsofuse .
