@cyanheads/art-institute-chicago-mcp-server
Search the Art Institute of Chicago collection: artworks, artists, exhibitions, and audio guides via MCP. STDIO or Streamable HTTP.
6 Tools • 1 Resource
Overview
The Art Institute of Chicago's collection: about 133,000 artworks, plus artists, exhibitions, and audio-guide stops. Search with filters and facet counts, read full records with provenance and IIIF image URLs, and search audio-guide transcripts. No API key required.
Tools
Resources
artic_get_artworks returns the same record for clients that don't surface resources.
Skills
The server serves it over MCP as skill://art-institute-chicago-guide/SKILL.md; it also ships in skills/ in the npm package and Docker image.
Capability reference
Searches and vocabulary lookup accept limits up to 1,000, and the response states the cap actually applied: 100 rows for artworks, artists, exhibitions, and the audio guide, and 1,000 values for vocabulary. Defaults are 10 rows (audio guide 2, vocabulary 25). Keep the same requested limit when following next_page; page times applied size must stay within 1,000. At the vocabulary cap, narrow with contains.
artic_search_artworks tool
query (every word must match, except common English words such as the and of, which are ignored; "exact phrase", -exclude, and a | b work) plus filters artist, artist_id, department, artwork_type, style, subject, classification, material, technique, theme, place_of_origin, gallery, year_from / year_to (date-span overlap, negative for BCE), acquired_from / acquired_to (fiscal year of acquisition, inclusive; works with no recorded year never match), public_domain_only, on_view_only, and has_image, combined with AND- Up to 100 rows per page (default 10), within the first 1,000 matches; rows carry
main_reference_number and, when recorded, acquisition_fiscal_year. sort is relevance, date_asc, date_desc, acquired_desc (recent acquisitions first), or acquired_asc; works missing the sorted field come last, and sort_applied reports popularity when relevance had no query text
facets adds the top 15 values for any of 11 fields (the ten vocabularies and artist; artist counts every credited artist and rows carry artist_id); limit: 0 returns counts only
artic_get_artworks tool
ids takes up to 1,000 entries (comma strings up to 100,000 characters); blanks and duplicates are dropped and artwork page URLs are read as ids. The first 100 are fetched; sections chooses from description and provenance (default), exhibition_history, publication_history, and catalogue, and is the size lever: every section on 100 well-documented works comes to about 1.4 MB- Records return whole, in request order.
missing_ids lists ids the museum doesn't have; deferred_ids lists ids past the first 100, to pass to the same tool again
include_related_media (default on) loads up to 100 linked lectures and audio stops; description_attribution appears whenever CC BY description text is returned
artic_search_artists tool
query (all name words must match) or ids, not both; ids takes up to 1,000 entries (comma strings up to 100,000 characters) and the first 100 are fetched. Query mode adds artists_only (default true), born_from / born_to, and up to 100 agents per page- Each agent carries
artwork_count, up to three sample_works (the museum's highlights first), life years, and alt_names; ids mode reports missing_ids and deferred_ids to pass as ids in the next call
ulan_id is the Getty ULAN identifier when the museum records one; coverage is sparse, and it is distinct from the museum id used by other tools
artic_search_exhibitions tool
query, when (current, upcoming, past, or the default any), date_from / date_to (YYYY-MM-DD, matched by run overlap), and exact artwork_id / artist_id filters, combined with AND; up to 100 per page within the first 1,000 matches. query ignores common English words such as the and of- Relationship filters cover recorded museum links; absent links do not prove an artwork or artist never appeared in a show
sort is relevance, start_desc, or start_asc, defaulting to relevance with a query and start_desc without; status is the museum's label and doesn't say whether a show is open (when does)- A missing date is unrecorded, not a run that did not happen: a show with no closing date is past when marked
Closed, else current from its opening, and a date window reads it as running on its opening date alone (Closed) or on from it; a show with only a closing date ran on that date alone, and one with no dates matches no window
- Rows carry dates, gallery, summary, web page, image,
artist_ids, and the artworks shown, when the museum lists them
artic_search_audio_guide tool
query is required and matches stop titles and transcripts (every word except common English words such as the and of, which are ignored); up to 100 whole stops per page (default 2)- Each stop has
title, audio_url (MP3), and transcript but no artwork id; every response carries license_text and source_citation (content is for noncommercial educational and personal use)
artic_lookup_vocabulary tool
vocabulary is one of department, artwork_type, style, subject, classification, material, technique, theme, place_of_origin, or gallery; optional contains substring (ignoring case and accents, so applique finds appliqué (technique)), public_domain_only, and up to 1,000 values (default 25)- Values come back most common first with
artwork_count, in the exact form the artic_search_artworks filter of the same name accepts; filter_param names that filter
artic://artworks/{id} resource
{ artwork, license_text, description_attribution?, notice? } as application/json, where artwork is the artic_get_artworks record with its default sections and related media- An unknown id fails as
artwork_not_found; ids come from artic_search_artworks
Features
Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.
Art Institute-specific:
- Keyless access to the Art Institute of Chicago public API (
api.artic.edu/api/v1); IIIF image URLs are built from each record (843 px for every image, 1686 px and a IIIF manifest for public-domain works), never fetched
- One shared request pacer under the API's published limit of 60 requests a minute, retries for transient failures inside a 20-second deadline, and an in-process response cache (records 6 hours, searches 15 minutes, vocabularies 24 hours)
- Text search requires every word to match, except common English words such as the and of, which are ignored, so totals are real and a miss reads as zero hits; results keep the museum's relevance order
- Placeholder years in the museum's data (outside −8000 to 2100) are left out of output, year filters, and date sorts;
date_display stays the authority. Date and acquisition sorts keep works with no value for the sorted field and list them last
Agent-friendly output:
- Rights travel with the data:
image.rights (public_domain / in_copyright) on every image, with the 1686 px URL only where reuse is allowed; the API's license_text verbatim; description_attribution when CC BY text is returned; source_citation on audio-guide results
- Partial results instead of failures:
artic_get_artworks reports missing_ids and deferred_ids, and when a secondary lookup (related media, artist counts) fails, the primary records still return with a notice naming what is missing
- Paging that names the next move:
totalCount, has_more, next_page, and a notice for the next page, the 1,000-match ceiling, or the filter to loosen after zero hits
Data and licensing
The Art Institute of Chicago licenses its API data by surface, and the server passes the API's own license_text through with every artwork, artist, exhibition, and audio-guide result:
This server is an independent project and is not affiliated with or endorsed by the Art Institute of Chicago.
Known limitations
- Only the first 1,000 matches of any search are reachable without an authenticated key. Broad questions need filters or facets.
- The API's limit of 60 requests a minute is per egress IP, so every client behind one IP shares it. A call that cannot start within its 20-second budget fails as
rate_limited.
- Curatorial text is sparse: about 10% of artworks have a
description, and in a general sample 94% lack provenance. About 1% of artists have a biography, and there is no nationality field, only artist_display prose.
- About 15 artworks carry placeholder years and about 4,900 carry no dates; neither matches a year filter.
- About a third of artworks (47,677 of 133,165, 2026-11-03) have no recorded acquisition fiscal year; they never match
acquired_from / acquired_to and sort last by acquisition year.
- Some vocabulary titles are stored cut at 40 characters in the museum's data (
gelatin silver (developing-out-paper) pr). Pass values as artic_lookup_vocabulary lists them.
- The API's firewall refuses any request whose text contains markup such as
<script>; the call fails as request_blocked.
- About 4% of exhibitions list their artworks. 650 of 6,262 have no recorded closing date (2026-11-03); the date filters read those as described above.
- Audio-guide stops have no artwork link. Some titles are file names, and some transcripts are in Spanish.
- Image URLs can stop resolving when the museum unpublishes or replaces an image, and relevance order follows the museum's ranking, which may shift.
Getting started
Connect to the hosted instance, no install:
Every caller of the hosted instance shares one api.artic.edu request budget of 50 requests a minute. For sustained use, run your own instance.
Or run it locally:
For other clients, add this to the MCP config file:
No API key or account is needed. The Art Institute API asks clients to identify themselves: set AIC_CONTACT to an email or URL the museum can reach (see Configuration). For Docker or a source build, see Running the server.
Configuration
See .env.example for every server setting and the common framework overrides.
Running the server
Local development
Requires Bun v1.4.0 or higher (or Node.js v24+).
-
Clone and install:
-
Build and run:
-
Run checks and tests:
Docker
The image defaults to HTTP transport, stateless session mode, and logs to /var/log/art-institute-chicago-mcp-server. To run it over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "-e", "[email protected]", "ghcr.io/cyanheads/art-institute-chicago-mcp-server:latest"]. To build locally, docker build -t art-institute-chicago-mcp-server .; OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.
Project structure
Development guide
See CLAUDE.md for development guidelines and architectural rules. The short version:
- Handlers throw, framework catches — no
try/catch in tool logic
- Use
ctx.log for logging, ctx.enrich for paging and notices
- Register new tools and resources in the barrels under
src/mcp-server/*/definitions/index.ts
- Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields
Contributing
Issues are welcome. Run checks and tests before submitting:
License
This project is licensed under the Apache 2.0 License. See the LICENSE file for details.
