Tool

Wikidata MCP

A provenance-first Wikidata MCP server. Every answer carries its receipt: the statement node, rank, references, retrieved_at, and backend — reproducible from primary source.

Tools

resolve
Ranked candidates for a term or property. Never auto-picks; returns all matches with confidence scores so you choose.
get_claims
Fetch statements from an entity, optionally filtered by property. Returns the full receipt: rank, references, retrieved timestamp, and backend.
entity_seo_pack
Query by name or QID. Returns sameAs identifiers, JSON-LD ready, plus conflicts and external-ID mappings. Useful for knowledge graph alignment.
query_sparql
Raw SPARQL queries against WDQS. Automatically wraps results with provenance metadata (retrieved_at, backend).

Resource: wikidata://sparql-cheatsheet for query help.

Install and run

Wikidata MCP runs as a standard Model Context Protocol server over stdio. 

pip install -e ".[dev]"
wikidata-mcp

Requires Python 3.8+. Full setup and integration instructions are in the GitHub repository.

Tests

Pure module tests run in isolation:

pytest -q

Add live smoke tests against WDQS (the real Wikidata Query Service): 

WIKIDATA_MCP_LIVE=1 pytest -q

What it can't claim

  • Freshness equals WDQS lag. retrieved_at is fetch time, not edit time. The server reflects what WDQS reports at query time, which may lag live edits by seconds to minutes.
  • External ID sameAs are confidence-reviewed, not resolved. entity_seo_pack returns external-ID mappings as confidence: review with no formatter-URL resolution (formatter-URL resolution is out of v1 scope).
  • No SPARQL repair, caching tiers, or local mirror. The server always queries WDQS directly. No query optimization, no local cache, no offline mode. By design — see the spec.

Source

Wikidata MCP is open source on GitHub: github.com/esbuilds/wikidata-mcp. File issues, read the spec, or fork and extend.