art-mcp

An MCP server for searching museum collections and viewing artwork images and metadata — right inside Claude (or any MCP client).

It works out of the box with seven sources that need no API key, and can pull in three more if you provide free keys.

Sources

Looking beyond Europe and the United States

The first four sources are all Western institutions, which meant non-Western art was only visible through a collecting museum's Asian or African department. The six added since address that directly, but the coverage is uneven in ways worth knowing up front:

• Asia is well served. vam and japansearch are both keyless, and smithsonian adds the National Museum of Asian Art.

• Africa runs mainly through smithsonian (unit_code:NMAfA) and vam. There is no Africa-based institutional API to point at — Digital Benin is the most significant African digital archive but publishes no public API, and the British Museum's SPARQL endpoint no longer resolves.

• South America has no museum API at all. Not one Latin American institution publishes a usable collections API. Brazil's federal museums run Tainacan as one WordPress site per museum with no central index and patchy images; Mexico's Mediateca INAH now redirects to a bare repository. So the region is covered by smithsonian (unit_code:NMAI, for Andean and Amazonian material) and by wikidata, which reaches Brazilian and Latin American painting through Wikimedia Commons. This is a real gap, not an oversight — worth revisiting as digitisation programmes there mature.

Why there is no Google Arts & Culture source

It has no public read API. The only developer-facing programme, Large Scale Data, is a pipeline for institutions to publish into the platform, and artsandculture.google.com/robots.txt explicitly disallows /api/*, so the internal endpoint is off-limits on both robots and terms-of-service grounds. Going direct to the institutions that publish their own open data — which is what most of the sources above are — gets the same collections without the problem.

A note on Smithsonian

One adapter covers 21 museums, so narrowing helps. The API accepts Solr-style field filters straight in the query string: unit_code:NMAfA (African Art), unit_code:FSG (Freer and Sackler, Asian art), unit_code:NMAI (American Indian).

Write the filter as a plain AND term — Nasca AND unit_code:NMAI. Wrapping free text and a field filter in parentheses makes EDAN drop the filter and match most of the institution.

DEMO_KEY works for a first try but is capped at 10 requests/hour.

NMAI is metadata-only in practice. Its records are flagged as having images, but the API returns no media for them — not in search results and not on the detail endpoint — which is consistent with the museum's cultural protocols around Indigenous material. So unit_code:NMAI is genuinely useful for finding and reading about Andean and Amazonian works, and get_artwork_image will almost always come up empty for them. NMAfA and FSG both serve images normally.

A note on Japan Search

Japan Search is the National Diet Library's cross-sector aggregator — national museums, university archives and prefectural collections behind one index. It catalogs in Japanese and has no English titles, so query it in Japanese (北斎, 浮世絵) for decent recall. Results come back with language: "ja" and the native title in titleOriginal; that is correct output, not a failure. It is also the only source that reports a real per-object licence, which lands in license.

A note on Wikidata

This is a fallback for regions no museum API reaches, not a curated catalogue. Metadata quality varies far more per item than a museum's own records, and a search resolves the keyword to an entity and then returns works by it — so it answers "Tarsila do Amaral" well and a subject phrase badly. Images come from Wikimedia Commons.

A note on Rijksmuseum

This uses Rijksmuseum Data Services, which needs no key — the older www.rijksmuseum.nl/api endpoint was retired and now answers 410 Gone.

The new API serves Linked Art and has no free-text search parameter, so search_artworks fans the keyword across title, creator, and description and interleaves the hits. Search also returns bare identifiers, so each result costs a follow-up lookup (plus two more to resolve its image) — expect Rijksmuseum to be the slowest source in an all-museum search.

A note on Art Institute of Chicago

Search and metadata work, but the IIIF image host sits behind a Cloudflare bot challenge, so get_artwork_image will usually fail for artic with a 403.

Related MCP server: artic-mcp

Tools

• list_sources — which museums are queryable right now.

• search_artworks { query, source?, limit? } — search one source or all of them; returns compact results with a source + id for each.

• get_artwork { source, id } — full metadata for one artwork.

• get_artwork_image { source, id } — downloads the image and returns it so the model can actually see it.

Connect to Claude

No install step — npx fetches the package on demand.

Claude Code

claude mcp add art-mcp -- npx -y art-mcp

Claude Desktop

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "art-mcp": {
      "command": "npx",
      "args": ["-y", "art-mcp"]
    }
  }
}

Restart the client, then try: "Search the Met for Van Gogh and show me an image."

Adding API keys

All optional — seven sources work without any key. The server reads keys from the process environment, so pass them through your MCP client's env block; a bare .env file is not loaded automatically.

{
  "mcpServers": {
    "art-mcp": {
      "command": "npx",
      "args": ["-y", "art-mcp"],
      "env": {
        "HARVARD_API_KEY": "your-key-here",
        "SMITHSONIAN_API_KEY": "your-key-here",
        "TEPAPA_API_KEY": "your-key-here"
      }
    }
  }
}

SMITHSONIAN_API_KEY is the one most worth setting: it is instant and self-service from api.data.gov, and it is what unlocks the African, Asian and Indigenous American collections.

Reliability

Museum APIs rot quietly. The Rijksmuseum endpoint this server was originally built on was retired and started answering 410 Gone with no announcement — the kind of break that only surfaces when a user hits it.

So a nightly workflow runs the full search → detail → image path against every live API and opens an issue when a source that used to work stops working. Run it yourself with:

npm run smoke

Known-broken checks (currently the Art Institute's Cloudflare-blocked images) are reported but don't fail the run, so a red badge always means something new actually broke.

Development

npm install
npm run build     # compile to dist/
npm run dev       # run from source with tsx
npm run inspect   # build + open the MCP Inspector
npm run smoke     # exercise every provider against the live APIs

License

MIT