@pipeworx/wikidata

Connect: https://gateway.pipeworx.io/wikidata/mcp · Install: one-click buttons

Tools: 5

Wikidata is the free, structured-data sister of Wikipedia. ~110 million items (people, places, works, concepts) with machine-readable properties and relationships. The closest thing the open web has to a universal knowledge graph. Free, no auth, supports SPARQL queries.

Why this matters for AI agents

For entity disambiguation, structured facts about anything notable, or graph traversal (“what books did this author write that won prizes?”), Wikidata is the canonical open source. Pair with Wikipedia (prose) and OpenAlex (academic specifics) for full open-knowledge coverage.

Common flows:

• Entity lookup. “Who is Paul McCartney?” → Q2599 with structured facts (birth date, nationality, occupation, band memberships).
• Property query. “What’s the population of France?” → Q142, property P1082 (population), with references and timestamps.
• Disambiguation. “Paris” → Q90 (capital of France), Q830149 (Paris, Texas), Q3296 (Paris, Greek mythology). Wikidata’s structured types resolve which “Paris” the agent is asking about.
• Graph traversal. SPARQL: “All Nobel laureates in Physics born in Italy” → joinable across structured properties.

Tools

Reading labels in another language

search_entities, get_entity, get_entities and get_wikidata_facts all take language. It is passed straight to the API’s own languages param, so a single code ("ka") or a comma/pipe-joined list ("ka,ru,en") both work, and a list returns every requested language’s label, description and aliases side by side in one upstream call — the shape you need to decide whether two records in different languages are the same entity. Omit it and you get English, exactly as before.

Two fields, and the difference matters:

• labels / descriptions / aliases_by_language are keyed by language and contain only what Wikidata actually has. Ask for a language the entity has no label in and the key is simply absent — that absence is the honest answer, and it is how you detect a translation gap.
• label / description are the single best value, and they fall back to English when the requested language is missing. So get_entity({id:"Q162887", language:"zu"}) returns label: "Enguri HPP" with an empty labels map: there is no Zulu label, and you can tell.

get_wikidata_facts localizes both halves — property names and their values — so {id:"Q162887", language:"ka"} comes back with Georgian keys (ქვეყანა → საქართველო), not English keys with Georgian values.

mul is a real language code here, and you will see it. Wikidata moved language-neutral labels — personal names, most Latin binomials, many place names — to the mul (“multiple languages”) code, and wbgetentities then returns an empty en label map for those items. Q42 (Douglas Adams) is one: it has an English description and no English label. The fallback chain is therefore requested → en → mul, and when the name came from mul you see it under a mul key in the labels map:

{"id":"Q42","label":"Douglas Adams","labels":{"mul":"Douglas Adams"},
 "description":"British science fiction writer and humorist (1952–2001)"}

That key is the provenance: the name is not attributed to a language the entity does not carry. Before this the pack returned label: null for Q42 — a clean 200 with a null where the answer was.

One bad id used to zero the batch

wbgetentities is inconsistent about unknown ids, and the two cases look nothing alike:

• An unknown but well-formed id (Q999999999) comes back marked missing, and every other id in the request still resolves.
• An id outside the item-id range or plainly malformed (Q99999999999, NOTAQID) makes the API refuse the entire request — HTTP 200, no entities at all, a top-level {"error":{"code":"no-such-entity","id":"Q99999999999"}}.

Read naively the second case is an empty map, so one typo in a 50-id batch reported all 50 as not_found: a clean 200 saying nothing resolved. The error names the offending id, so get_entities drops it and retries (bounded, one id per pass). A batch of ["Q162887","Q99999999999","Q90","NOTAQID","Q1326165"] now resolves the three real entities and returns the two bad ids in not_found.

Pinning statements to an edit

Wikidata is live and continuously edited, so “the current statements” is only meaningful relative to a revision. Both entity tools return lastrevid, modified and a revision_url, so an answer built from them can be pinned to the exact edit it was read at, and re-checked later for drift.

wikidata_recent_changes answers the other half of that: which items changed, and when. Namespace 0 is items (Q-ids), 120 is properties (P-ids), 146 is lexemes. type is new (freshly created), edit (edits to existing items) or all. Use since for a time window and cursor to page further back.

Auth

None. Wikidata is fully open. Wikidata’s SPARQL endpoint has fair-use rate limits but generous for most agent traffic.

Identifier scheme

Every entity has a stable Q-number that serves as a permanent identifier across Wikipedia language editions. Embed Q-numbers in agent output as canonical citations.

Common pitfalls

• Quality varies wildly by entity. A famous person (Einstein, Q937) has hundreds of well-sourced properties. A small-town politician may have 5 properties, half from auto-imported sources. Always check the references field on properties you’re acting on.
• Multiple values for “current” properties. “President of company X” may have 8 historical values plus a current one. Wikidata uses qualifier properties (start time, end time, “preferred rank”) to indicate which is current — but agents often grab the first value naively.
• Property duplication. Multiple properties can express related concepts (“country” P17, “country of origin” P495, “country of citizenship” P27). Picking the right one matters for accuracy.
• Vandalism risk. Wikidata is editable like Wikipedia. High-profile entities are watched, but obscure ones can carry vandalism for days. Don’t quote a single Wikidata fact as ground truth without a references chain to an authoritative source.
• Translation gaps. Labels exist in many languages but not always in the one you asked for. label falls back to English, then to the language-neutral mul label, when that happens, so a non-English request never comes back empty — but it also means a returned label is not proof the entity HAS a label in that language. Read the labels map for that: a missing key is a real gap, and a mul key means the name is language-neutral rather than translated. See “Reading labels in another language” above.
• SPARQL timeouts. Complex graph queries can hit the SPARQL endpoint’s 60-second timeout. Decompose into smaller queries or use property-specific lookups.
• Identifiers, not facts. Wikidata is best for “what’s the canonical ID of this entity” and “what does it link to.” For deep biographical or historical narrative, follow the linked Wikipedia article.

Data sources

• MediaWiki Action API — https://www.wikidata.org/w/api.php
  • action=wbsearchentities (search), action=wbgetentities (entity read, props=info|labels|descriptions|aliases|claims|sitelinks, languages= for the language filter, up to 50 pipe-joined ids per call)
  • action=query&list=recentchanges (recent changes feed)
• API docs: https://www.wikidata.org/w/api.php and https://www.mediawiki.org/wiki/API:Recentchanges
• Entity pages: https://www.wikidata.org/wiki/<QID>; a specific revision: https://www.wikidata.org/w/index.php?oldid=<revid>

All endpoints are keyless. Wikidata asks callers to send a descriptive User-Agent; this pack sends Pipeworx-Wikidata-MCP/1.0.

Tools

• search_entities — Search Wikidata entities by label or alias (e.g., “Albert Einstein”, “Python programming language”, “Tokyo”). Returns entity IDs, labels, descriptions, and aliases. Useful for finding the Wikidata ID
• get_entity — Get full Wikidata entity by ID (e.g., “Q42” for Douglas Adams, “Q5” for human, “Q1764” for Budapest). Returns labels, descriptions, aliases, claims/statements (properties and values), sitelinks, and t
• get_entities — Batch version of get_entity — resolve up to ${MAX_BATCH_IDS} Wikidata Q-ids in ONE call instead of looping single get_entity calls. Same fields as get_entity per entity (labels, descriptions, aliases,
• get_wikidata_facts — Structured facts about a Wikidata entity in HUMAN-READABLE form — property names and values resolved to labels, not raw P/Q codes. PREFER OVER get_entity for “what is X\x27s ”, “facts about
• wikidata_recent_changes — Newest items created or edited on Wikidata right now, from the live MediaWiki recent-changes feed. Answers “which Wikidata item was created most recently”, “what QIDs were created today”, “what change