@pokemontcgapi/mcp
[npm] [license] [Glama]
An MCP server for the Pokémon TCG API at
pokemontcgapi.com. It gives an agent thirteen tools over the whole
catalogue: international, Japanese and Simplified Chinese print lines, card names in
eight locales, illustrators, images, and prices that carry their source, basis, grade and sample
size. The current counts are live at
/v1/status and broken down on
coverage.json.
Unofficial. Not produced, endorsed, supported by or affiliated with Nintendo, Creatures Inc.,
GAME FREAK inc. or The Pokémon Company International. Pokémon and all related marks are trademarks of
their respective owners.
Get a key
Generate the Idempotency-Key once per signup and keep it with the request body:
Lost the response? Repeat the exact same request (same Idempotency-Key, same body byte for byte, same network: same public IPv4 or the same IPv6 /64) within 24 hours and the response comes back, if stored, secret included; it is the original response, so a key rotated or revoked since then is not revived. A new Idempotency-Key for the same email returns 409 ACCOUNT_EXISTS; the same key with a different body returns 409 IDEMPOTENCY_CONFLICT.
We store only a hash of the key; the signup response is kept for 24 hours so the same request can be replayed. Save data.key.secret now.
If replay is unavailable, sign in and rotate the key, or use /v1/accounts/recover with an already verified email to get a new secret.
The key comes back in data.key.secret. Confirming the address we email raises the trial from
80 to 800 credits, and the trial ends 30 days after signup. Paid plans start at 29 EUR a month:
pricing.
Install
Claude Code:
Claude Desktop — claude_desktop_config.json:
Cursor — .cursor/mcp.json:
VS Code — .vscode/mcp.json. Note the top-level key is servers, not mcpServers, and inputs
keeps the key out of the committed file:
Environment: PTCG_API_KEY, needed by twelve of the thirteen tools, and PTCG_BASE_URL (defaults to
https://api.pokemontcgapi.com). Node ≥ 20. The exception is ptcg_get_reference, which reads a
public route. ptcg_get_catalogue_status is not an exception: it starts on the public /v1/status
and then reads one set per print region, which needs the key. Without it the server starts and lists
its tools, then those twelve calls come back asking for it.
The tools
Thirteen tools, not one per endpoint. tools/list sits in the model's context on every turn, so the
whole surface is about 13 KB, and each tool is shaped like a question rather than like a route —
the model does not have to chain four calls to answer one thing.
Every tool is annotated readOnlyHint: true and destructiveHint: false. Nothing here writes.
ptcg_identify_card_from_image and the three centering tools are marked idempotentHint: false:
each invocation consumes credits. Do not retry paid work automatically.
Centering is included on Growth and higher; Developer is excluded. A verified Free trial includes
5 measure and 5 confirm calls, independently. Grade and standards have no separate trial call counter.
Measure accepts base64 image bytes, explicit front/back face, optional standard/small card_size and
corrected corners (TL, TR, BR, BL in EXIF-oriented pixels reduced to longest edge <=2560 px).
It returns ratios, uncertainty and editable geometry; needs-review with null ceilings is expected.
After correcting corners with another measure, a person checks or moves cut and print lines in the
app. Grade ceilings arrive after that app calls confirm, or from known ratios through grade.
There is deliberately no confirm tool. An assistant cannot assert that a person reviewed the
lines. Only the app collecting that person's explicit acceptance may send confirmed_by: person.
Always display ceiling from centering alone, not a prediction of the final grade. Never label a
result verified or certified. Keep source status, incomplete flags and uncertainty visible.
See the two-step guide.
Commercial refusals put next_step.handoff on the first line of the tool result, followed by the API message and complete details. Show that sentence and its URL to the account owner verbatim and do not retry. The owner completes checkout, email verification or the contact step. CARD_ALLOWANCE_EXCEEDED is one of them: the period has retrieved its distinct-card allowance (1,000 for the trial, 30,000 a month on Developer); cards already retrieved keep answering, new ones need the next period or Growth.
Importing a card catalogue
For a complete card import, use the REST API directly:
GET /v1/cards?limit=250&orderBy=id, then follow links.next verbatim. The flat list fills pages
across set boundaries and uses fewer requests than a separate card loop for every set.
Add include=translations for names at the plain catalogue cost. Priced includes have separate tariffs.
For a single print region, add q=set.region:JP or q=set.region:CN; lang only selects a name translation.
Use /v1/sets/{code}/cards when you need a particular set and /v1/sets?region=JP to browse set metadata.
The quickstart contains both
paging loops and dated measurements. The
migration guide explains how to capture
a change feed watermark before importing and maintain the replica afterwards.
The MCP search tool is intended for bounded interactive searches; its region argument is sent to
the API as set.region:JP (or CN, WEST), so a Japanese search reads Japanese rows only. Use REST
for a complete import.
What this API does not have
The last tool exists because of this section, and it returns these facts from a live call rather
than leaving a model to infer them:
- No Korean cards. Zero
KR sets and zero ko translations. The print region and the locale are
modelled in the schema and carry no data, so filtering on them returns an empty result, not an error.
- Card game text is English, and uneven.
attacks, abilities, weaknesses, resistances,
subtypes, retreat_cost, rules and flavor_text carry rows since 3 September 2026, on the
20,725 Western printings. Measured on 16 September 2026 against 57,450 cards: attacks on 29.9% of
the whole catalogue and 82.9% of the Western part, subtypes 35.0%, weaknesses 28.0%,
flavor_text 17.9%, abilities 7.0%, rules 5.1%. Japanese and Chinese printings carry none, so a
null attacks means we do not hold it, never that the card has no attack.
- No format legalities. The card object carries no
legalities field, and level is empty. If the
question is about deck legality, this API cannot answer it.
All three are measured, dated in the source, and repeated verbatim in the tool descriptions, so an
agent is told before it calls rather than after.
Reading prices correctly
There is no printing filter. First Edition, Unlimited, holofoil, reverse holofoil and graded rows all
come back together, so read printing, condition and grading on each row rather than taking the
first number. basis separates GUIDE (published upstream) from DERIVED (computed by us);
PTCG_INDEX is our own composite in EUR and carries sample_n. Every observation has an as_of
date and is delayed by at least a day — never quote a price without it.
What the plan withholds is named rather than hidden: graded and non_english_locales for a trial
key, graded on Developer, nothing from Growth up. The API says so in meta.withheld on the prices
route, in the X-Plan-Withheld header when prices ride on a card, and in a top-level withheld
field on the batch. So a card with no graded rows may be the plan speaking, not the catalogue.
For ptcg_get_cards, the API's missing field is authoritative when present, including its
suggested_id values. The tool keeps its existing missing array of strings and adds
missing_details, an array of { id, suggested_id? }; the text also shows each suggestion.
If the API omits missing, the tool falls back to comparing requested ids with returned id
and legacy_id, ignoring case and repeated ids. This supports older API versions without
another request, but cannot discover their unreported canonical-set collisions or suggestions.
The current API omits missing when every id resolves, so that fallback then returns an empty array.
ptcg_get_card_prices reads a card with its prices, so the exclusions arrive in that header.
Sealed products
/v1/sealed carries booster boxes, booster packs, elite trainer boxes, tins, blisters and
collections across the international, Japanese and Simplified Chinese lines. Each product carries
languages, the languages it physically exists in; the Simplified Chinese line, 273 products added
on 30 October 2026 in the cn- sets, carries ["zh"]. The API has no language filter on that list,
so ptcg_search_sealed with language scans pages of 100 at 2 credits each, keeps the rows whose
languages contain it, and says when it stopped before the end of the catalogue. Without language
it is one page, 1 credit per 50 rows, cursor included.
ptcg_get_sealed_prices reads the product (1 credit) and then its quotes (2 credits). Sealed rows
carry no condition, printing or grading; locale is the language the box is printed in, and the
Chinese line carries Cardmarket LOW and TREND under zh. index is the catalogue price of the
product in euro with its date, the same index_eur as on the product row, with sample_n null and
by_locale empty, because the composite is computed for cards only. Rows under a locale other than
en follow the plan rules of the cards, and the tool names them in the same withheld line.
Context discipline
Results are capped at 50 rows regardless of what the API allows, sent as aligned tables rather than
JSON, with a compact field projection. A table is shorter than the same rows as JSON because the keys
are not repeated on every row; we do not publish a percentage, because we have no reproducible
measurement to show next to it.
Truncation is always announced along with the cursor to continue. Price rows are the one thing never
truncated.
Protocol
Built on @modelcontextprotocol/server v2, which negotiates the
2025-11-25 revision and accepts clients
down to 2024-10-07. stdio transport.
The revision is the library's, not a claim of our own: SUPPORTED_PROTOCOL_VERSIONS in
@modelcontextprotocol/[email protected] tops out at 2025-11-25, so a client that asks for anything
newer is answered with that. Verified against the published package, not read off a changelog.
Also available
Build from source
Node >= 20. npm test runs the unit tests in tests/. What CI enforces is that the package
typechecks and builds on both Node 20 and Node 22, and that npm pack produces
the file list the registry is meant to receive.
This package is developed inside the private monorepo that runs
pokemontcgapi.com and mirrored here on each release,
so a merged pull request travels back by hand rather than by merge button. That
is not a reason to send patches elsewhere — open the issue or the PR here, it is
the address that gets read.
Licence
MIT. Data served by the API carries per-source redistribution terms — see
https://pokemontcgapi.com/legal/attribution.
