Skip to content

revanbcn/wikidata-google-knowledge-mcp

v0.2.1MIT

Token-efficient MCP for Wikidata and Google Knowledge Graph search, entity resolution, relationships and deterministic identity reconciliation.

Changelog

All notable changes to this project are documented here. The project follows Semantic Versioning; before 1.0 minor versions may change behaviour.

Unreleased

Changed

  • README and documentation link the official MCP Registry entry (server.json name, PyPI package, stdio).

[0.2.1] - 2026-09-28

Release-safety patch. Package code, resolver policy and outputs are unchanged.

Fixed

  • PEP 740 attestations are uploaded to PyPI: uv publish only matches attestations among the paths it is given, and 0.2.0 was uploaded without them. The release job now passes the attestation files, checks offline that every distribution has one before uploading, and fails unless PyPI shows provenance for every file afterwards.
  • The public-tree check keeps its list of private names only as hashes, so the check itself no longer names them, and in CI it rejects a commit that carries a personal email address.

Changed

  • Documentation and README install from PyPI first; the GitLab release tag stays as an alternative. The release check verifies PyPI version pins.
  • The documentation build has git, so sitemap.xml carries last-modified dates.
  • Classifiers: console environment, science/research audience, artificial intelligence topic.

[0.2.0] - 2026-09-28

Discovery and first-use release. Resolver policy and decision records are unchanged.

Added

  • Documentation site built from this repository (scripts/build_site.py) and published with GitLab Pages: installation for Claude Code, Cursor and Codex, a tool reference generated from the MCP tool schemas, entity resolution, batch and evidence, benchmark and troubleshooting.
  • Reproducible offline benchmark (benchmarks/replay_benchmark.py) replaying captured Wikidata MCP responses, and a recorded live demo with its recorder script and text transcript.
  • server.json for the official MCP Registry (PyPI package, stdio) and the matching mcp-name ownership marker in the README. The registry entry is published only after the package is live on PyPI and the namespace's domain proof is in place.
  • wdkg --version; wikidata-google-knowledge-mcp --help and --version, which answer without starting the stdio protocol. Unknown arguments are reported on stderr, never on stdout.
  • Release pipeline: build once, verify the wheel and sdist in clean environments, publish the same files to PyPI with Trusted Publishing and attestations, attach checksums to the GitLab release.

Changed

  • wdkg --help and every subcommand's help now carry examples, a plain explanation of AUTO_MATCH, HOLD, AMBIGUOUS, CONFLICT and NO_CANDIDATE, and the exit codes; every resolve flag has a description.
  • --pretty and --no-cache are accepted after the subcommand as well as before it.
  • The missing Google key message says where to put the key and that Wikidata works without it; wdkg status adds a setup hint for Google when no key is present. The error code (not_configured) is unchanged.
  • Package metadata: documentation URL, source URL, keywords; plugin manifests point to the documentation site.

[0.1.0] - 2026-09-27

First public release.

Added

  • Stdio MCP server wikidata-google-knowledge-mcp with five read-only tools: kg_search, kg_entity, kg_related, kg_resolve and kg_status.
  • wdkg CLI over the same core: search, entity, related, sparql (guarded), status, cache, resolve, resolve-batch, export-evidence, validate-evidence.
  • Compact outputs: at most 5 candidates (3 by default), selected properties, byte-budgeted responses, bounded relationships and hierarchy.
  • Deterministic entity resolution with closed-vocabulary evidence and conflict codes, kind rules for places, organizations, people, events, event series and works, and explicit HOLD / AMBIGUOUS / NO_CANDIDATE results.
  • Optional Google Knowledge Graph support with exact joins to Wikidata P646 (/m/) and P2671 (/g/), kept separate from local identity evidence.
  • Resumable JSONL batch resolver with request budgets, cache reuse, per-row errors and an auditable evidence export.
  • Portable Agent Skill and plugin manifests for Claude Code, Codex, Cursor and the Agent Plugins format.