whitney-museum-mcp-server
This server lets you search and retrieve records from the Whitney Museum's public API through MCP, covering artworks, artists, exhibitions, events, and audio guides.
Search artworks by title, artist, medium, classification, date, on-view status, department, or random sort; get slim records with a first image URL.
Get full artwork details by ID — dimensions, credit line, description, all images, alt text, and attribution rights/citation.
Search artists by name, collection/Biennial/on-view filters; retrieve artist bios, Getty ULAN and Wikidata IDs.
List an artist's artworks or exhibitions using the artist's record ID for accurate matching.
Search exhibitions (from 1931) and events (from 2008) with title filters, date ranges, and raw Ransack predicates.
Get individual exhibition records and search audio guides (from 2009).
Run raw queries against any endpoint (artists, artworks, exhibitions, events, guides, pages) with arbitrary Ransack filters and sorting.
Return results as readable markdown or structured JSON, with pagination control (1–30 per page) and source attribution included.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@whitney-museum-mcp-server search for artworks by Alexander Calder"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
whitney-mcp-ts
An MCP server for the Whitney Museum's public API. Search the collection, artists, exhibitions, events and audio guides from Claude Desktop or any MCP client.
TypeScript, stdio only, friendly parameters mapped onto the API's Ransack syntax, travels with attribution, and responses trimmed to keep them manageable in a model's context.
The Whitney's API is open and unauthenticated; no API key needed.
Install
npx whitney-mcp-tsOr from source:
git clone https://github.com/mildlydiverting/whitney-mcp-ts.git
cd whitney-mcp-ts
npm install
npm run buildNode 20 or newer.
Related MCP server: Wikipedia MCP Server
Claude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"whitney": {
"command": "npx",
"args": ["-y", "whitney-mcp-ts"]
}
}
}From source, point at the build instead:
{
"mcpServers": {
"whitney": {
"command": "node",
"args": ["/absolute/path/to/whitney-mcp-ts/dist/index.js"]
}
}
}Quit and reopen Claude Desktop. The config doesn't expand ~, so use a full path.
Tools
Tool | What it does |
| Search 27,000+ works by title, artist, classification, medium, date, on-view status |
| Full record for one work — dimensions, credit line, description, image URLs |
| Search 7,000+ artists; filter by collection, Biennial, on view |
| Full artist record, including biography, Getty ULAN and Wikidata IDs |
| Works by a given artist |
| Exhibitions featuring a given artist |
| Exhibitions from 1931 onwards, with date filtering |
| One exhibition record |
| Talks, tours and programmes from 2008 onwards |
| Audio guides published since 2009 |
| Any endpoint, with raw Ransack predicates |
What to expect
Search returns slim records.
Whitney records carry long HTML descriptions and biographies, so search gives you the basics and whitney_get_* gives you everything. Prose is truncated at 2,000 characters, whole responses at 25,000.
Pagination is 30 per page.
That's the API's page size. page picks the page, limit trims it further — the default of 10 stops a broad search flooding your context.
Search syntax is Ransack
Predicates nest under q (_eq, _cont, _cont_all_split, _gteq and friends), sorting under q[s]. The typed tools map friendly arguments onto these. whitney_query exposes them raw, for anything the typed tools don't reach.
Artist and artwork fields are mapped explicitly.
Exhibitions, events, guides and pages go through a generic summariser instead: HTML stripped, timestamps shortened to dates, relative URLs made absolute, internal foreign keys dropped. The Museum says its field set may change, so nothing is hard-coded that doesn't need to be. Run a search with limit: 1 to see the real field names, then use filters for anything specific.
Attribution
Every response ends with a source line naming the Museum and linking their terms. Detail records also carry two attribution fields:
rights— the copyright notice, extracted from the Whitney's own description textcitation— a ready-made markdown line, eg:
Edward Hopper, Early Sunday Morning, 1930. Whitney Museum of American Art, © Heirs of Josephine N. Hopper/Licensed by Artists Rights Society (ARS), New York.
Note the rightsholder is usually not the artist. In a sample of ten works,
nine differed — estates, foundations, and rights societies like ARS and VAGA.
So rights is worth reading rather than assuming; it's the notice the
Whitney's terms ask you to keep.
Works with no copyright notice — older material, mostly — get a citation ending after the Museum link.
Bulk Data
For anything at scale, don't page through the API. The Whitney publishes artist and artwork CSVs at https://github.com/whitneymuseum/open-access, released under CC0.
Known API quirks
sort: "random"doesn't reliably randomise. Asking for three random works returned three consecutive accession numbers by the same artist. If you need genuine randomness, request a random page number instead.display_dateon artworks is free text — "1915–1931, printed 1976–1977" — so date ranges don't work there. Exhibitions and events have real timestamps, andstarts_on_or_after/starts_on_or_beforefilter on those.Living artists have
death_year: "0". Don't do arithmetic on it.Ransack ignores predicates it doesn't recognise rather than erroring, so a misspelled field silently returns everything. Check your result counts.
outputSchemawas dropped from ten of the eleven tools (all butwhitney_query). The MCP SDK (v1.30.0) converts a Zod v3 output shape to JSON Schema via a vendoredzod-to-json-schemathat always stamps$schema: draft-07, which some MCP clients reject outright ("unsupported dialect" — they validateoutputSchemaagainst 2020-12 only). Structured content is still returned in every response; it's just no longer pre-advertised in the tool list.
Security
The Whitney's API needs no key, so there's no credentials, nothing to leak and nothing to configure.
Every tool is read-only. The server makes GET requests to whitney.org and nothing else — no filesystem access, no shell, no writes. Endpoint names come from a fixed list, record IDs are URL-encoded, and filter parameters go through URLSearchParams.
One known limitation: the HTML-to-markdown conversion uses regular expressions with nested quantifiers, which could backtrack badly on deliberately malformed input. The input comes from the Whitney's own CMS rather than from users, so if this happens, something has gone very wrong in NYC and it's probably not my fault.
If you find something, open an issue.
Development
npm test # unit tests, offline
npm run test:canary # live checks against the Whitney API
npm run inspect # MCP Inspector against the built serverThe canary suite runs monthly in CI. It exists to catch the Whitney changing a field name — the failure mode that unit tests can't see, since they run against saved fixtures.
Licences
Code: MIT License.
The Whitney provides the API with the following note:
The data accessed through this API may be protected by copyright, and other restrictions, of the Whitney Museum of American Art and third parties. You may use this data for noncommercial educational and personal use and for "fair use" as authorized under law, provided that you also retain all copyright and other proprietary notices contained on the materials and cite the author and source of the materials.
Read their terms at https://whitney.org/about/website/api before you build anything public on it.
Note the CSV datasets linked above are CC0, which may be more permissive than the API terms.
Not affiliated with the Whitney Museum of American Art. This is a third-party wrapper around their public API, which is documented at https://whitney.org/about/website/api.
Thanks
See also Sam Parsons' whitney-museum-mcp — a Python/FastMCP server over the same API, with Docker and HTTP transport.
To the Whitney's digital team for publishing an API at all, and for documenting it properly. And for writing nice blog posts that explicitly say they use claude code, because that makes me feel less guilty for vibe coding this whole shebang.
Available Tools
11 toolswhitney_artist_artworksList an artist's artworksARead-onlyIdempotent
List the collection works by one artist.
Args:
id (string): artist ID
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[] } with slim artwork records.
Use this rather than whitney_search_artworks with an artist name — it matches on the artist record rather than the credit line, so collaborative and attributed works are handled correctly.
Inputs: id (required), page, limit, response_format. Returns: note, page, count, total, +4 more
whitney_artist_exhibitionsList an artist's exhibitionsARead-onlyIdempotent
List Whitney exhibitions featuring one artist.
Args:
id (string): artist ID
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[] }.
Exhibition fields are passed through generically (HTML stripped, long prose truncated), because the Museum documents these fields as subject to change. Coverage is fuller for recent decades than for the Museum's early history.
Inputs: id (required), page, limit, response_format. Returns: note, page, count, total, +4 more
whitney_get_artistGet a Whitney artistARead-onlyIdempotent
Fetch one artist's full record, including their biography where the Museum has published one.
Args:
id (string): TMS ID or 'T'-prefixed reference
response_format ('markdown' | 'json', default 'markdown')
Returns: { record: { id, name, sort_name, dates, birth_year, death_year, in_collection, in_biennial, in_artport, on_view, ulan_id, wikidata_id, biography, page_url } }
The biography is stripped of HTML and truncated. ulan_id (Getty ULAN) and wikidata_id are useful for cross-referencing other sources.
Inputs: id (required), response_format. Returns: record, source
whitney_get_artworkGet a Whitney artworkARead-onlyIdempotent
Fetch one artwork's full record by ID.
Args:
id (string): the artwork's TMS ID, as returned by whitney_search_artworks
response_format ('markdown' | 'json', default 'markdown')
Returns: { record: { id, title, artist, artist_ids[], date, medium, dimensions, classification, department, accession_number, credit_line, on_view, description, object_label, visual_description, alt_text, images[], page_url } }
Fields are omitted when empty. Alt text is human-written where available and AI-generated otherwise; alt_text_is_ai_generated flags which.
Errors: returns a 404 message if the ID does not exist.
Inputs: id (required), response_format. Returns: record, source
whitney_get_exhibitionGet a Whitney exhibitionARead-onlyIdempotent
Fetch one exhibition record by ID.
Args:
id (string): exhibition ID, as returned by whitney_search_exhibitions
response_format ('markdown' | 'json', default 'markdown')
Returns: { record: {...} } — fields as published by the API, HTML stripped.
Inputs: id (required), response_format. Returns: record, source
whitney_queryRaw Whitney API queryARead-onlyIdempotent
Escape hatch for queries the typed tools do not cover. Hits any endpoint with arbitrary Ransack predicates.
Args:
endpoint ('artists' | 'artworks' | 'exhibitions' | 'events' | 'guides' | 'pages')
id (string, optional): fetch a single record instead of a list
filters (object, optional): raw predicates, e.g. { "medium_cont_all_split": "graphite paper" }
sort (string, optional): e.g. 'popularity desc'
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: a list result, or { record: {...} } when 'id' is supplied.
Prefer the typed tools where they fit — they return tidier records. Use this to discover field names, combine unusual predicates, or reach the 'pages' endpoint. Note that 'pages' mirrors the Museum's CMS structure and is rarely useful outside it.
Inputs: endpoint (required), id, page, sort, +3 more.
whitney_search_artistsSearch Whitney artistsARead-onlyIdempotent
Search the 7,000+ artists listed by the Whitney — those with work in the collection and those who have appeared in Museum exhibitions.
Returns slim records (id, name, dates, in_collection, in_biennial, on_view). Call whitney_get_artist for the biography and authority IDs.
Args:
name (string, optional): substring match on the display name
in_collection, in_biennial, on_view (boolean, optional)
sort ('default' | 'name_asc' | 'name_desc' | 'most_viewed' | 'random')
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[], note? }
Example: "which O'Keeffes are in the collection?" -> search name="O'Keeffe", then whitney_artist_artworks with the returned id.
Inputs: name, page, sort, limit, +4 more. Returns: note, page, count, total, +4 more
whitney_search_artworksSearch Whitney artworksARead-onlyIdempotent
Search the Whitney Museum's online collection of 27,000+ works.
Returns slim records (id, title, artist, date, medium, classification, on view, first image URL). Call whitney_get_artwork for the full record including dimensions, credit line, description and all images.
Args:
title, artist, medium, date (string, optional): substring matches
classification (string, optional): exact match, e.g. "Drawings"
department ('collection' | 'artport' | 'special', optional)
on_view (boolean, optional)
sort ('default' | 'title_asc' | 'title_desc' | 'most_viewed' | 'random')
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[], note? }
Examples:
"watercolours with flowers in the title" -> title="flower", classification="Drawings", medium="watercolor"
"show me something at random from the collection" -> sort="random", limit=3
Notes:
Medium spellings follow US usage ("watercolor", "color").
Date is free text ("1915-1931, printed 1976-1977"), so filter loosely and check results.
Inputs: date, page, sort, limit, +7 more. Returns: note, page, count, total, +4 more
whitney_search_eventsSearch Whitney eventsARead-onlyIdempotent
Search Museum events — talks, tours, public programmes and so on — going back to 2008. This list is largely comprehensive.
Args:
title (string, optional): substring match
filters (object, optional): raw Ransack predicates
sort (string, optional): e.g. 'start_time desc'
starts_on_or_after / starts_on_or_before (string, optional): ISO dates (YYYY-MM-DD) filtering on start_time
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[] }
As with exhibitions, fields are passed through generically. Sort by 'start_time desc' for the most recent programming.
Inputs: page, sort, limit, title, +4 more. Returns: note, page, count, total, +4 more
whitney_search_exhibitionsSearch Whitney exhibitionsARead-onlyIdempotent
Search exhibitions mounted by the Whitney, going back to 1931. Records are fuller for recent decades and sparse for the Museum's early history.
Args:
title (string, optional): substring match
filters (object, optional): raw Ransack predicates for fields this tool does not expose
sort (string, optional): e.g. 'start_time desc', 'random'
starts_on_or_after / starts_on_or_before (string, optional): ISO dates (YYYY-MM-DD) filtering on start_time
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[] }
Exhibition fields are passed through generically — the Museum documents its field set as subject to change, so this tool keeps whatever scalar fields the API returns, strips HTML, shortens timestamps to dates and drops internal IDs. Known fields include title, start_time, end_time, date_override, url, primary_text, press_highlights and popularity. Run one search with limit=1 to confirm before writing a 'filters' query.
Inputs: page, sort, limit, title, +4 more. Returns: note, page, count, total, +4 more
whitney_search_guidesSearch Whitney audio guidesARead-onlyIdempotent
Search audio guides published since 2009. Fetching an individual guide also returns its stops.
Args:
title (string, optional), filters (object, optional), sort (string, optional)
page (number, default 1), limit (number, 1-30, default 10)
response_format ('markdown' | 'json', default 'markdown')
Returns: { total, count, page, has_more, next_page?, results[] }
Inputs: page, sort, limit, title, +2 more. Returns: note, page, count, total, +4 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
Each resource has a clear search/get or relationship tool, and the overlap between whitney_artist_artworks and whitney_search_artworks is explicitly resolved in the descriptions. whitney_query is positioned as an escape hatch for unsupported queries, not as a competing primary tool.
The whitney_ prefix with search_/get_ resource names forms a clear pattern for most tools. The exceptions are whitney_artist_artworks and whitney_artist_exhibitions, which use a possessive noun style rather than a verb_noun style, and whitney_query, which is a one-off.
11 tools is well-scoped for a museum API covering artists, artworks, exhibitions, events, and guides, plus relationship helpers and a query escape hatch. Each tool has a distinct job and none feel redundant.
Core read-only workflows are covered: search and detail retrieval exist for artworks, artists, and exhibitions, and searches exist for events and guides. The main gap is the lack of typed getters for individual events and guides, which the guide description implies should exist; whitney_query can work around this but it is less discoverable.
Maintenance
Related MCP Connectors
Art Institute of Chicago MCP — wraps the ARTIC public API (free, no auth)
Victoria and Albert Museum (V&A) Collections MCP. Keyless. Wraps the V&A Collections API v2.
Unlock a world of art with the Met Museum MCP! Use the 'search_artworks' API to find stunning
Search and manage your team's knowledge from MCP clients.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceInteract with the Cooper Hewitt API with MCP.2MIT
- AlicenseNot gradedqualityDmaintenanceProvides structured access to Wikipedia content including search, summaries, images, links, and more via MCP tools.6Apache 2.0
- AlicenseNot gradedqualityCmaintenanceEnables searching and retrieving details from the Harvard Art Museums collection.79 npm2MIT
- AlicenseNot gradedqualityCmaintenanceEnables searching the Metropolitan Museum of Art collection by subject, artist, or keyword, retrieving detailed artwork information, and listing curatorial departments via natural language.93 npm3MIT