rijksmuseum-mcp+
# rijksmuseum-mcp+
[](https://modelcontextprotocol.io/specification/2026-07-28)
[](https://www.npmjs.com/package/@modelcontextprotocol/server)
[](https://github.com/modelcontextprotocol/ext-apps)
## Overview
The **rijksmuseum-mcp+** MCP server lets you explore the Rijksmuseum's artwork collections through natural conversation with an AI assistant. It does this by creating a [bridge](https://www.anthropic.com/news/model-context-protocol) between the AI system's chat environment and an enriched copy of the museum's [open-access, curated metadata](https://data.rijksmuseum.nl). This in turn enables [many features beyond those offered by the Rijksmuseum](/docs/compared-to-rijksmuseum.md) through their [Search API](https://data.rijksmuseum.nl) and [collections portal](https://www.rijksmuseum.nl/en/collection/). Rijksmuseum-mcp+ works best when used together with [rijksmuseum-iconclass-mcp](https://github.com/kintopp/rijksmuseum-iconclass-mcp), a sibling resource for searching and exploring [Iconclass](https://iconclass.org) concepts.
> Please do not treat the data made available by this resource as current or authoritative. It is based on data copied from the Rijksmuseum on May 2nd, 2026. For current data, please always use the Rijksmuseum's own [search portal](https://www.rijksmuseum.nl/en/collection/) and [APIs](https://data.rijksmuseum.nl). Nor have the (in small part, LLM based) enrichments of the museum's provenance data been reviewed or endorsed by the Rijksmuseum. This is a technology demo that remains – by its nature – in active development.
This MCP server was developed by the [Research and Infrastructure Support](https://rise.unibas.ch/en/) (RISE) group at the University of Basel. We are particularly interested in exploring the research opportunities, methodological risks, and technical challenges posed by retrieving and analysing data with LLMs. If you are interested in collaborating with us in this area, please [get in touch](mailto:rise@unibas.ch).
## Key Features
- **Searching artwork metadata**. You can search artworks by keyword, by structured filters (artist, type, material, technique, date, physical dimensions, production place), or by meaning — a semantic search that handles interpretive queries like "melancholy winter scenes at dusk". There's also [iconographic search via Iconclass codes](https://kintopp.github.io/rijksmuseum-mcp-plus/iconclass-visualization), so you can ask for works depicting a specific scene or motif rather than just matching words in titles.
- ✨ **Searching digitised artworks visually**. A multimodal search can match a plain-language description such as "children playing in the street" against the digitised images themselves, not only against catalogue metadata. Each result then notes whether it was found through the image, the catalogue text or the Iconclass codes. The search can be further narrowed with filters, for example to paintings, a period, or a single artist.
- **Looking closely at works.** Any digitised artwork can be opened in an interactive, inline viewer. You can instruct the AI-assistant to inspect image regions by itself — useful for analysing a specific area of the image you've highlighted for it in the viewer. If an artwork has curator defined related artworks (e.g. preparatory sketches, or different impressions of the same design) these can be accessed through the viewer as well — its < / > buttons step through them in place, without leaving the viewer.
<p align="center"><img src="docs/van-der-Ast.jpg" alt="Stilleven met vruchten en bloemen (SK-A-2152) with highlighted detail containing three shells" width="600"></p>
- **Collection-level analysis**. You can ask for statistical breakdowns across the whole collection: top creators, distributions by decade, type, or theme, geographic spread, as well as demographic questions.
- **Provenance and ownership history**. Trace who owned a work and when, which works passed through a particular collector or dealer, sales and confiscations in a given city or period, price histories, and how long families held their collections. This is made possible by an [experimental AAM parser](https://kintopp.github.io/rijksmuseum-mcp-plus/provenance-parser-visualization.html) that permits structured, [CMOA/PLOD-aligned queries](https://kintopp.github.io/rijksmuseum-mcp-plus/provenance-patterns.html).
- **Scholarly apparatus**. Lookup bibliographies for individual works, reverse lookups (which artworks cite a given publication), and conservation histories including technical examinations such as X-rays, infrared, and dendrochronology.
- **Relationships between works**. A similarity engine ("find images similar to..") [compares works across multiple dimensions](https://kintopp.github.io/rijksmuseum-mcp-plus/similar-to-SK-A-1115.html) — visual, thematic, lineage, shared subject — and surfaces pendants, pairs, copies, reproductive prints after paintings, and different impressions of one design.
- **People and places**. You can search persons by profession, lifespan, or birthplace and then view their works, or run geospatial queries (e.g. "works depicting places within 20 km of Haarlem").
- **Linked Open Data**. Works carry persistent handle.net URIs and other external IDs, and entities (creators, materials, depicted persons and places, themes) carry identifiers linking them to Wikidata, VIAF, ULAN, and RKD.
- **Command-line interface.** The [bundled `rijks-mcp` tool](https://github.com/kintopp/rijksmuseum-mcp-plus/blob/main/docs/technical-guide.md#cli) runs the same queries from the terminal — each tool exposed as a verb, with JSONL output for piping into tools such as `jq` so that results are scriptable and reproducible.
## Sample Queries
The system is designed to let you search, explore and ask questions about the Rijksmuseum's collections in natural language. For example:
*What is the actual distribution of Rembrandt's works in the Rijksmuseum across painting, printmaking, and drawing — and how does this challenge popular perceptions of him as primarily a painter?* [Demo](https://claude.ai/share/e2070805-94a5-4e35-9aea-24b6274b641d)
For examples of more complex queries and sample responses, please browse the [research scenarios](docs/research-scenarios.md). Nine scenarios for art historians and curators cover iconographic, semantic and multimodal search, image inspection, similarity search, provenance, collectors' marks, bibliography and conservation records, each with a shared demo transcript.
## Quick Start
The best way to get started is with [Claude Desktop](https://claude.com/download) or [claude.ai](https://claude.ai) by adding rijksmuseum-mcp+ to Claude as a *remotely hosted*, [custom 'Connector'](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) using the URL below. This is currently free for one connector – additional connectors require a paid ('Pro') or higher [subscription](https://claude.com/pricing) from Anthropic.
```
https://rijksmuseum-mcp-plus-production.up.railway.app/mcp
```
Go to _Customize_ → _Connectors_ → _Add custom connector_ → Name it as you like and paste the URL into the _Remote MCP Server URL_ field. You can ignore the Authentication section. Once the connector is configured, optionally set the permissions for its tools (e.g. 'Always allow'). See Anthropic's [instructions](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) for more detailed instructions.
Many other desktop and web-based clients such as OpenAI's ChatGPT or Mistral's Vibe and open-source applications support remotely hosted, custom MCP servers. Please consult their documentation for more information. In particular, the free version of [Mistral's Vibe](https://chat.mistral.ai) already has excellent support for MCP servers and skills (see below). In contrast, ChatGPT can currently (October, 2026) only support the full functionality of this server (in particular, viewing and interacting with images in the chat timeline) in conjunction with a paid Business or Enterprise subscription. So at the moment the best way to use the hosted MCP server for free is either with Claude or with Mistral. Alternatively, you can install rijksmuseum-mcp+ locally on your own computer. Please consult the [technical guide](/docs/technical-guide.md) for more details.
Afterwards, follow the same procedure to install rijksmuseum-mcp+'s sibling resource for IconClass, [rijksmuseum-iconclass-mcp](https://github.com/kintopp/rijksmuseum-iconclass-mcp). This allows you to automatically search and explore c. 1.3 million [IconClass](https://iconclass.org) notations, concepts, and descriptive texts alongside the Rijksmuseum's metadata.
## Research skill
The `rijksmuseum-mcp+` skill file ([rijksmuseum-mcp-plus.skill](skill/rijksmuseum-mcp-plus.skill)) gives the AI assistant detailed guidance in natural language on how to use rijksmuseum-mcp+ effectively: which tool to choose for a given question type, how to combine searches, important metadata distinctions and known limitations. The package also includes reference files with full description of the available provenance search patterns and the `find_similar` functionality. Skills were [originally developed by Anthropic](https://support.claude.com/en/articles/12512176-what-are-skills) for their Claude products but have since become an [open standard](https://agentskills.io/home). Making use of this skill is optional but will significantly improve the quality and efficiency of your AI assistant's responses when exploring the collection. The downloaded skill file can be installed in Claude by following [these instructions](https://claude.com/resources/tutorials/teach-claude-your-way-of-working-using-skills).
Currently (October, 2026) ChatGPT's free tier makes uploading skill files cumbersome. Instead, use [Mistral Vibe](https://mistral.ai/products/vibe/) and install the skill via the 'folder' option, first downloading and unpacking [`rijksmuseum-mcp-plus.zip`](skill/rijksmuseum-mcp-plus.zip).
## Technical Design
For local setup (stdio or HTTP, also via cli), deployment, architecture, data sources, and configuration, please see the [technical guide](/docs/technical-guide.md). For measurements of what the server costs and saves in tokens compared with plain web search, see the (still early and experimental) [benchmarks](/docs/benchmarks.md).
The enriched metadata is available as a standalone SQLite database ([`vocabulary.db.gz`](https://github.com/kintopp/rijksmuseum-mcp-plus/releases/download/v0.90/vocabulary.db.gz), from the [v0.90 release](https://github.com/kintopp/rijksmuseum-mcp-plus/releases/tag/v0.90)). Its tables, joins and example queries are described in the [database schema](/docs/database-schema.md).
## Roadmap
Ongoing:
- fix bugs and fine-tune queries and tool descriptions
Later:
- paper/presentation
Maybe:
- OCR/HTR of appropriate images in the collection
- incorporate historical exhibition data
- integrate with other Linked Open Data resources (e.g. [Colonial Collections](https://data.colonialcollections.nl))
- support inferred geolocation data
- improve the `description` signal for *find_similar* (e.g. via a LLM re-ranker)
## Authors
[Arno Bosse](https://orcid.org/0000-0003-3681-1289) — [RISE](https://rise.unibas.ch/), University of Basel with [Claude Code](https://claude.com/product/claude-code), Anthropic.
## Citation
If you use rijksmuseum-mcp+ in your research, please cite it as follows:
**APA (7th ed.)**
> Bosse, A. (2026). *rijksmuseum-mcp+* (Version 0.96) [Software]. Research and Infrastructure Support (RISE), University of Basel. https://github.com/kintopp/rijksmuseum-mcp-plus
**BibTeX**
```bibtex
@software{bosse_2026_rijksmuseum_mcp_plus,
author = {Bosse, Arno},
title = {{rijksmuseum-mcp+}},
year = {2026},
version = {0.96},
publisher = {Research and Infrastructure Support (RISE), University of Basel},
url = {https://github.com/kintopp/rijksmuseum-mcp-plus},
orcid = {0000-0003-3681-1289},
note = {Developed with Claude Code (Anthropic, \url{https://www.anthropic.com})}
}
```
## Image and Data Credits
Collection data and images are provided by the **[Rijksmuseum, Amsterdam](https://www.rijksmuseum.nl/)** via their [Linked Open Data APIs](https://data.rijksmuseum.nl/).
**Licensing:** Information and data that are no longer (or never were) protected by copyright carry the **Public Domain Mark** and/or **[CC0 1.0](https://creativecommons.org/publicdomain/zero/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](https://creativecommons.org/licenses/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](https://www.rijksmuseum.nl/en/rijksstudio).
**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](https://data.rijksmuseum.nl/policy/information-and-data-policy) for the full terms.
**Enriched data:** The enrichments this project adds to the vocabulary database, such as the parsed provenance tables, and the embeddings database (`gemma.db`) are released under **[CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/)**. The Rijksmuseum records they are built on keep the licences described above.
> This project was inspired by [@r-huijts/rijksmuseum-mcp](https://github.com/r-huijts/rijksmuseum-mcp), the original Rijksmuseum MCP server based on the museum's now superseded REST API.
## License
The software in this project is licensed under the [MIT License](LICENSE).
TDQS
Scored across 13 tools
Tools are largely distinct and descriptions include explicit 'Not for X — use Y' clauses; the four viewer-related tools (get_artwork_image, inspect_artwork_image, navigate_viewer, poll_viewer_commands/remount_viewer) are clearly differentiated, with internal ones labeled. However, several descriptions redirect agents to tools that aren't in this set (semantic_search, search_persons, search_provenance, collection_stats, search_inscriptions), which creates ambiguous dead-end pointers.
Nearly all tools follow a verb_noun pattern (search_artwork, get_artwork_details, list_curated_sets, browse_set, inspect_artwork_image, navigate_viewer). Minor deviations exist: singular 'artwork' vs plural 'artworks' across search_artwork / find_artworks_citing_publication, and a couple of noun-first phrasings, but the set reads predictably.
13 tools is well within the sweet spot and each earns its place: core discovery, detail retrieval, image interaction, bibliography, conservation, curated sets, and delta sync are all represented without redundancy. Internal viewer tools are small helpers that don't bloat the surface.
Core lifecycle for browsing the collection (search → details → image → bibliography/conservation → sets → recent changes) is covered. But descriptions repeatedly point to missing tools (semantic_search for concept queries, search_persons for demographic vocabIds, collection_stats for aggregates, search_inscriptions), leaving genuine dead ends an agent cannot work around cleanly.