rijksmuseum-mcp+ lets you explore the Rijksmuseum's artwork collections through natural conversation with an AI assistant. It does this by creating a bridge between the AI system's chat environment and the museum's open-access, curated metadata.
This project was inspired by @r-huijts/rijksmuseum-mcp, the original Rijksmuseum MCP server based on the museum's now superseded REST API.
rijksmuseum-mcp+ was developed at the Research and Infrastructure Support (RISE) group at the University of Basel and builds on our ongoing work on benchmarking and optimizing humanities research tasks carried out by large language models (LLMs). We are particularly interested in exploring the research opportunities and technical challenges posed by using structured data with LLMs for humanities research. If you are an interested in collaborating on these questions, please get in touch.
You can explore artworks using rijksmuseum-mcp+ with the same (with minor exceptions) search filters offered by the Rijksmuseum on their search collections page. Beyond this, it provides the following additional features:
-
Full-text corpora — (
description,inscription,provenance,creditLine,curatorialNarrative). This permits, for example, comparative analyses of the collection's catalogue (description) and curated wall texts (curatedNarrative). -
Semantic search — this enables multilingual, concept/meaning-based explorations across multiple metadata categories. For example, queries like "vanitas symbolism" or "sense of loneliness in domestic interiors" which can't be expressed as structured metadata.
-
Spatial dimensions — proximity radius searches on the museum's (
nearPlace) and size filters (minWidth/maxHeight) enable spatial queries like "artworks related to places within 25 km of Leiden" or "prints smaller than 10 cm wide" as well as two new parameters (nearLatandnearLon) to enable spatial queries from arbitrary locations. -
More metadata — several more metadata fields not not accessible from the museum's search portal (
birthPlace/deathPlace,profession,iconclass, andcollectionSet). -
Iconclass — access to its own Iconclass database, cross-linked with the Rijksmuseum's metadata, which can be searched and explored not just by notation by also title, description, parent/child classes and semantically by concept.
-
Interactive Image Viewer — view high-resolution images of artworks inline in your chat discussion (this feature requires Claude Desktop or claude.ai). Zoom, pan, rotate, flip horizontally or view the image full-screen.
-
Image analysis (experimental) — the AI assistant can analyse images visually in combination with its own background knowledge and the artwork's structured data (e.g. "which iconographic elements of the Annunciation in this image have corresponding entries in Iconclass?"). It can also annotate the image with these elements in the image viewer (e.g. "identify the biblical scenes depicted in the painting's panels and highlight these for me").
-
Structured outputs — As most of the data provided by rijksmuseum-mcp+ is in structured form, it's often straightforward for the AI assistant to also represent or export these in a structured manner (e.g. tabular formats) or drawn on them for follow-up tasks, such as visualizations or other analyses.
The best way to get started is with Claude Desktop or claude.ai by adding a custom 'Connector' to Claude using the URL below. This currently requires a paid ('Pro') subscription from Anthropic.
https://rijksmuseum-mcp-plus-production.up.railway.app/mcp
Goto Settings → Connectors → Add custom connector → Name it whatever you like and paste the URL shown above into the Remote MCP Server URL field. Once the connector is configured, set the permissions for its tools (e.g. 'Always allow'). See Anthropic's instructions for more details.
rijksmuseum-mcp+ can also be run as a local MCP server in STDIO mode with local copies of its metadata and embedding databases. Please see the technical notes for details.
Technically speaking, rijksmuseum-mcp+ is a Model Context Protocol (MCP) server. As such, it also works with many other browser based chatbots including those whose large language models (LLMs) can be used without a paid subscription. Mistral's LeChat is a good example. It's also compatible with many open-source desktop 'LLM client' applications such as Jan.ai that are able to make use of local or cloud based LLMs, and agentic coding tools such as Claude Code or OpenAI Codex.
In comparison, OpenAI's ChatGPT still only offers limited, 'developer mode' support for MCP servers and while Google has announced MCP support for Gemini it has not indicated when this will be ready.
After you've added the rijksmuseum-mcp+ 'connector' (aka custom MCP server) to your AI system, test that everything is working correctly by asking your AI assistant to confirm its access: "Which MCP tools can you use to explore the Rijksmuseum's collections?". It should show you a list like this.
After that, ask your own questions:
- What artworks evoke vanitas and mortality?
- Show me artworks depicting places near the Oude Kerk in Amsterdam
- What works have the iconclass code for fabulous animals?
- Which artworks have a provenance linked to Napoleon?
- What are the 'top ten' works in the Rijksmuseum?
- I'm looking for artworks with inscriptions mentioning 'luctor et emergo'
- Show me sculptures in the collection by artists born in Leiden
- Which paintings are wider than 3 meters?
- Has anyone taken a tumble in SK-A-1718? Show me where.
For examples of more complex questions, see the research scenarios.
to be added Please see this reference for a comprehensive overview of all available search parameters.
to be added For now, please see this file or consult the DeepWiki entry for this repo. The rijksmuseum-se repo provides Jupyter notebooks and standalone scripts as starting points for interactive explorations of UMAP and PacMAP visualizations of the vocabulary embeddings.
Say what you are actually looking for, not how to find it. The assistant generally does better when given a research question than a list of parameters. "What prints were made after paintings by Rembrandt?" works better than "search for prints with technique etching by Rembrandt", because the first framing lets the assistant choose the right combination of tools and strategies.
For broad queries, add a second constraint. Searches across very broad categories — all paintings, all works on paper, all portraits — can match tens of thousands of records. The results are not ranked by importance and only a sample is returned (see 'Known Limitations' below). Combining with a date range, production place, or second subject term makes the results both faster and more meaningful.
Specify "paintings" when that is what you want, especially for concept searches. Paintings are underrepresented in concept-based (semantic) search results relative to prints and drawings, because those have denser subject tagging in the Rijksmuseum's catalogue. Saying "paintings showing X" rather than just "artworks showing X" helps the assistant apply the right corrections.
Try a concept search when structured filters return nothing useful. If searching by subject, Iconclass, or description doesn't find what you're looking for, asking the assistant to try a concept search (semantic search) can find artworks by meaning rather than exact vocabulary terms. This is especially useful for atmospheric or thematic queries. The assistant can also search Iconclass concept — finding the right notation code by meaning rather than exact keyword — and then use that notation for a precise structured search.
The MCP server (rijksmuseum-mcp+) seems stuck. If the server is not responding, it could be that it has been updated and the connection needs to be refreshed. To fix this, in your AI system's settings (e.g. in Settings in Claude Desktop or claude.ai) disconnect and reconnect the server, and then click on Configure to verify that all permissions are still correct. In other MCP clients, you may not be able to disconnect/reconnect. In that case, remove and add the server again.
Structured search results are not ranked by relevance. When filtering by subject, material, place, technique, or other structured fields, results currently come back in internal catalogue order — not by quality, importance, or closeness to the query. For a large result set, the first page is essentially an arbitrary slice of the matching artworks, not a curated selection. Concept-based (semantic) searches are the exception: those results are ranked by similarity to your query.
Result sets are capped and only partially paginated. Each search returns up to 25 results by default (up to 100 on request). When a query matches thousands of artworks, only a small, non-representative sample is returned. Adding more specific filters is the best way to get meaningful results from large collections and helps prevent the LLM being overwhelmed with metadata from too many search results.
Text coverage varies by field. About 61% of records include a cataloguer's description (in Dutch). Curatorial wall texts (in English) cover only about 14,000 artworks — mostly highlights and recent acquisitions.
Geolocation coverage is partial. About 64% of named production places have been geocoded. Proximity searches (e.g. "artworks produced near Delft") will miss artworks from places that haven't been geocoded. Where coordinates exist, they may point to the nearest town or region rather than a specific address.
Iconclass subject classification can be counterintuitive. The Iconclass system assigns subjects to specific branches of a strict hierarchy that does not always match everyday expectations. However, the assistant can search Iconclass by concept as well as by keyword — describing what you're looking for in plain language (e.g. "domestic animals" or "religious suffering") will often find the right notation even when the exact vocabulary term is unknown.
The collection data is predominantly in Dutch. Titles and subject tags are available in Dutch for virtually all records; English is available for roughly a third. The assistant will try both languages automatically, but searches for specialist terminology, historical place names, or older material may miss records that are catalogued only in Dutch.
Image analysis works better than image annotation. Anthropic's models are more accurate at describing the contents of an image than annotating it. For example, the models will often correctly describe they can 'see' (even drawing on the content rich description data for guidance) but struggle to draw accurate bounding-boxes around this content.
Soon:
- add support for
assigned_by attributionmetadata (e.g. "attributed to", "workshop of", and "follower of") - reduce dependencies on Rijskuseum Search API
- improve documentation
- fine-tuning
Later:
- address inconsistent pagination and ranked results issues
- add support for optional SKILL files
- add support for LLM analysis of user defined image selections
- address print vs. painting bias in semantic search from unbalanced subject labels
- review toponyms without clear geolocation data
- investigate support for MCP elicitations
- paper/presentation
Maybe:
- investigate integration with other Linked Open Data resources (e.g. Colonial Collections)
- investigate support for image similarity search (whole image, image segments)
Arno Bosse — RISE, University of Basel with Claude Code, Anthropic.
Collection data and images are provided by the Rijksmuseum, Amsterdam via their Linked Open Data APIs.
Licensing: Information and data that are no longer (or never were) protected by copyright carry the Public Domain Mark and/or CC0 1.0. Where the Rijksmuseum holds copyright, it generally waives its rights under CC0 1.0; in cases where it does exercise copyright, materials are made available under CC BY 4.0. Materials under third-party copyright without express permission are not made available as open data. Individual licence designations appear on the collection website.
Attribution: The Rijksmuseum considers it good practice to provide attribution and/or source citation via a credit line and data citation, regardless of the licence applied.
Please see the Rijksmuseum's information and data policy for the full terms.
This project is licensed under the MIT License.
