Data Packages MCP

scripts/data-packages-mcp/ is a read-only MCP server for assessing import quality from the outside: it reads only the published @museumwnf/*-data npm packages, the same files the public websites install. It never reads the inventory database or an exporter’s local output, and never publishes.

With it, an AI assistant such as Claude Code can:

• list every published data package and its versions, discovered from the npm registry;
• select a package version and browse it: items merged with their translations, the collection tree, partners, glossary, timelines, any record;
• run JMESPath queries over any file of any package family;
• read import-quality statistics: translation coverage, empty fields, items without images, partner item counts;
• compare two published versions, from a per-file summary down to one record field by field;
• view the images a record references.

Every record is shown with its inventory id and its legacy backward_compatibility key, and every lookup accepts either, so a finding leads straight to the record in /admin or to its legacy source.

Using it

The repository’s .mcp.json registers the server as mwnf-data-packages. It runs in Docker through the data-packages-mcp compose service; open the repository in Claude Code and approve the project server when asked. Start with list_packages, then select, then diff_versions or the browsing tools.

The tools, the package families, how downloads are verified and how to run the tests are documented in the project’s README.