rijksmuseum-mcp-plus

Rijksmuseum MCP+ Research Skill

Where the data comes from

The server answers from a snapshot of the Rijksmuseum's open data harvested in May 2026, not from the museum's live APIs. Records the museum has added, corrected or withdrawn since then are not reflected; get_recent_changes reads the live change feed, and the collection website is the current reference.

• OAI-PMH (https://data.rijksmuseum.nl/oai): one EDM metadata record per object, the backbone of the harvest.
• Linked Art (https://id.rijksmuseum.nl/): JSON-LD object records and vocabulary terms.
• Data dumps (https://data.rijksmuseum.nl/dumps/): person, group, place and concept authorities, the source of creator demographics and of links to Wikidata, VIAF, ULAN and RKD.

The museum publishes these under its information and data policy: CC0 1.0 or the Public Domain Mark, with CC BY 4.0 where it exercises its own copyright; it asks for a credit line either way. Each object's license field gives the rights for the work and its image: Public Domain Mark for most, In Copyright for roughly one work in eight, CC0 for a few.

This project enriched parts of the museum's data; the Rijksmuseum has not reviewed or endorsed any of it:

• Provenance texts and credit lines: parsed by a grammar into events, parties and ownership periods aligned with the AAM/CMOA standard. A few hundred ambiguous decisions out of many thousands were made by an LLM; these carry categoryMethod/positionMethod/correctionMethod flags, the model's reasoning, and a review page.
• Inscription field: split at query time, by rules, into individual marks, collector's marks and transcribed text.
• Place authorities: coordinates from Getty TGN (ODC-By 1.0), Wikidata (CC0) and GeoNames (CC BY 4.0), added only where a museum-supplied authority ID resolves them.
• Production records: each maker paired with their own role and attribution qualifier, which is what makes same-row attribution filters possible.
• Titles, descriptions, inscriptions and curatorial narratives: encoded together as one embedding per work, which semantic_search and multimodal_search's metadata channel both rank. The Dutch descriptions are also encoded on their own for find_similar's Description channel.
• Images and Iconclass notations: the photographs and the labels of the Iconclass notations used in the collection, encoded into the same image–text embedding space, so multimodal_search can rank a text query against what the pictures show. Subject terms are not part of the text either search matches.