The 30-second answer
Museums solved this problem once before, a century and a half ago, with the card catalog: a standard way for any visitor to ask a vast collection a precise question and get a precise answer. The Model Context Protocol — MCP — is the card catalog for AI assistants. It is an open standard that lets an assistant such as Claude reach out of its conversation and query a live, structured source of truth instead of relying on whatever it half-remembers from training. Mus3ums now publishes exactly such a catalog: an MCP server at https://direct.Mus3ums.com/mcp/ with seven tools. Five of them search the collection — artworks, artists, museums, articles, or all four at once. One returns an exact, real-money price quotation for a reproduction of any artwork in the catalog, as an art print or a hand-painted oil, at the size and in the currency we name. And one — the only tool that asks who we are — lets us check our own orders. No key, no registration, no payment is needed for any of the public tools. We connect the assistant once, and from then on "find me Caillebotte's rainy Paris street, and tell me what a hand-painted copy a metre wide would cost in euros" becomes a question with an exact answer.
1. What MCP is, without the jargon
An AI assistant, left to itself, is a scholar locked in a library that stopped acquiring books on the day its training ended. Ask it about Caravaggio's biography and it does well; ask it what a 70 × 90 cm hand-painted reproduction of The Calling of Saint Matthew costs today, in Canadian dollars, and it can only guess — politely, fluently, and unverifiably.
MCP, an open standard introduced by Anthropic in late 2024 and since adopted across the industry, fixes this in the simplest possible way: it gives the assistant a telephone. A website that has something worth asking about — a catalog, a database, a price list — publishes a small set of named tools, each with a description of what it does and what it needs to be told. The assistant reads that list the way we read a menu, and when our question calls for it, it dials: one structured request out, one structured answer back, quoted to us with the numbers intact. The assistant never rummages through the site's servers; it can only ask the questions the site has chosen to answer. That constraint is the point. A card catalog does not hand visitors the keys to the archive — it answers the query on the card.
2. The seven tools, and what each one answers
The server exposes seven tools. Notice the shape of the set: five are ways of finding things, one turns a find into a price, and one — the only private one — concerns us rather than the collection.
| Tool | What it answers | Token needed? |
|---|---|---|
search_artworks |
Full-text search of the artwork catalog by title, artist and description — "wheat field with cypresses", "Bruegel winter hunters". Each match returns the artist, the title, a link to the work's page, and a relevance score. | No |
search_artists |
Artists by name or biography, narrowed if we wish by birth city, birth country, or a year the artist was alive — a quiet superpower for questions like "painters active in Delft in 1660". | No |
search_museums |
Museums by name, optionally by city and country, with each result placed on the map: name, city, country, link. | No |
search_articles |
The sites' own essays, guides and FAQs, optionally filtered by category (blog, buyer's guide, support, and so on). | No |
search_fts |
All four catalogs at once — up to ten results from each, merged into a single list ordered by relevance. The right first move when we do not yet know whether "Vermeer Delft" should lead to a painting, a biography, or a museum. | No |
get_a_quotation |
An exact price for a reproduction of a named artwork — as an art print or a hand-painted oil painting — at a given width and height (25–250 cm per side, in cm or inches) and in a given currency. The artwork can be named by its catalog number or simply by title and artist. | No |
get_my_orders_details |
Our own orders — and only our own: dates, amounts, line items with their options, current status, shipping address. "Where is the Monet I ordered in March?" becomes a question our assistant can answer. | Yes — personal token from /profile/mcp/ |
Every search result carries a relevance score, and the tools return only good-confidence matches rather than padding the list — an assistant reading the results can tell a firm identification from a shrug. Results come back in the language we ask in: the language parameter accepts any of the sites' interface languages, and links point to pages in that language.
3. The quotation tool deserves a closer look
Price is where most "AI meets shopping" integrations go quietly wrong, so it is worth saying precisely what this one does. The quotation is computed by the same code that prices the buy page itself, and rounded the same way — the figure the assistant reads out is the figure we would see on the website, not an estimate of it. Three further habits of the tool repay attention:
- It refuses to guess about money. If we ask for a currency whose exchange rate is unavailable, the tool says so and stops. It never quietly substitutes US dollars and hopes we will not notice.
- It tells us what is and is not for sale. Not every work is offered in every form. Ask for a hand-painted oil of a work available only as a print, and the answer names the problem — and mentions the print, if that avenue is open.
- It holds the sizes honest. Each dimension must fall between 25 and 250 cm (about 10 to 98 inches) — the range a workshop can actually produce — and the tool converts inches for us rather than trusting anyone's mental arithmetic.
So a conversation like this becomes routine: "Find Gustave Caillebotte's Paris street scene in the rain, then quote me a hand-painted oil at 100 × 70 cm in euros, and the same size as a print, and tell me the difference." The assistant makes one search call and two quotation calls, and the comparison it hands back is arithmetic on real prices — the same comparison our guide to giclée prints versus hand-painted oils walks through by hand.
4. Connecting an assistant, in practice
The server speaks stateless MCP over plain HTTPS — JSON-RPC in a POST body, nothing more exotic — which means two audiences can use it.
With an MCP client
Any MCP-capable client connects with the endpoint URL alone. In Claude, that means adding a custom connector pointed at https://direct.Mus3ums.com/mcp/. In Claude Code, one command:
claude mcp add --transport http Mus3ums https://direct.Mus3ums.com/mcp/From that moment the seven tools appear in the assistant's repertoire and it decides, question by question, when to use them — no further configuration, and for the public tools, no credentials.
As a plain API
Developers who want the catalog without an MCP client can call the same endpoint directly. One curl command searches the collection:
curl -s https://direct.Mus3ums.com/mcp/ \
-H 'Content-Type: application/json' \
-d '{
"jsonrpc": "2.0", "id": 1, "method": "tools/call",
"params": {
"name": "search_artworks",
"arguments": {"query": "Caillebotte Paris street rainy day"}
}
}'The answer is structured JSON — catalog numbers, titles, artists, absolute URLs, scores — ready for whatever we are building.
The personal token, for orders only
To let an assistant see our orders, we sign in on the website, open Profile → MCP / API (/profile/mcp/), and copy the personal token shown there, sending it as an Authorization header. The design is deliberately narrow: the token identifies one account, the orders tool answers only for that account, and the search and quotation tools ignore the token entirely because they have nothing private to protect. Asked about an order number that is not ours, the server gives the same answer it would give for an order that does not exist — it declines to confirm even that much about anyone else's purchases.
5. Why a catalog site publishes an API at all
Because the questions have already moved. A growing share of people planning a purchase now begin not with a search engine but with an assistant — "what would a museum-quality copy of the Mona Lisa cost me?" is typed into a chat window, and the assistant answers from whatever sources it can actually verify. A site that publishes its catalog and its prices in a form assistants can query is simply doing what museums did when they printed their first public catalogs: meeting the visitor where the visitor already is, and preferring exact answers to folklore. The alternative — assistants guessing at prices and availability from stale training data — serves nobody, least of all the person about to spend real money on the answer. We have written elsewhere about how reproduction mediums and pricing actually work and what quality and value mean in a reproduction; the MCP server is those same answers, made machine-readable.
Three questions worth asking before connecting
Do we need an account or an API key to use the Mus3ums MCP server?
Not for most of it. The five search tools and the quotation tool are public and need no token, because they expose only what the websites already publish to every visitor. A personal token is required for exactly one tool — the one that returns our own orders, with a name, telephone and shipping address in the answer. That token comes from the /profile/mcp/ page after signing in, and it can only ever unlock the orders of the account it belongs to.
Is the price the assistant quotes the real price?
Yes. The quotation is produced by the same pricing code as the buy page and rounded identically, so the number the assistant reads out matches the number on the website. It covers prints and hand-painted oils from 25 to 250 cm per side, in any supported currency — and when a currency's exchange rate is unavailable, the tool says so rather than silently falling back to dollars.
Which assistants can connect?
Any client that speaks MCP over HTTP: Claude via custom connectors, Claude Code with a single claude mcp add command, and the widening field of MCP-capable desktop and developer tools. And because the transport is ordinary JSON-RPC over HTTPS POST, no MCP client is strictly required — a curl command or a few lines in any programming language reach the same seven tools.
Related on Mus3ums.com
- What the quotation tool is pricing: giclée prints versus hand-painted oil masterpieces, compared honestly.
- How reproduction prices are actually built from medium and commissioning choices: decoding art reproduction — mediums, pricing and commissioning.
- The search tools in human form: browse the artist catalog the same way the API does.
- Signed in already? The token and live tool list are on your profile's MCP / API page.
