The museum was founded in 1903 to be open to anyone who wanted to draw, study, copy and remake what it held. This API is that invitation in a digital form.
This URI content-negotiates: request it with
Accept: application/ld+json to be redirected to the catalog.
An index is not the collection. It is a way in — a list of handles that lets you reach for one thing among many without looking at everything first. Every index is a decision about what is worth finding, and that decision is made long before anyone searches.
Museums have always indexed. The card catalogue, the register, the thesaurus: each one fixes a small set of properties — maker, material, date, type — and makes those properties findable at the cost of everything it leaves out. What a period chose to record tells you what that period thought a collection was for. The gaps are as legible as the entries.
Those indexes were made by hand, so they had to be economical. A cataloguer could not record the colour of ten thousand objects, or how much of each object that colour covered, and no one asked them to. The limit was labour, and the categories grew around it.
Machines lift that limit, and in doing so they make new kinds of index possible: properties nobody catalogued because cataloguing them was never practical. This is not a better index, and it does not replace the ones made by hand. It measures something different, with its own blind spots — a colour index describes photographs as much as it describes objects, and it knows nothing of what any of these things mean. What it offers is another way in, and a set of questions the existing categories were not shaped to ask.
What follows are indexes of that second kind, built from the collection data and open for anyone to query, reuse or disagree with.
One caveat throughout. These indexes describe the objects published through this API — around ten thousand of the twenty-four thousand the museum holds. Which objects have been published, and in what order, is itself a decision made by people, and it shapes every figure below. This is a view of the catalogue, not of the collection.
Colours are extracted from the digital images: the background is removed, the remaining pixels are clustered, and each cluster is matched to a named tone and a broad base colour. What you see below is how much of the collection's visible surface each colour accounts for.
Twelve broad categories. Every colour the tagger finds is assigned to one of twelve base names — red, orange, brown, grey and so on — and the band shows how much of the collection's visible surface each accounts for. It is deliberately coarse: it answers "how much of this collection is blue" without pretending the answer is precise. Click a segment to see the tones it is made of, and the objects it dominates.
The same measurements, ungrouped. Each colour cluster is matched to the nearest name from an extended colour list — several hundred of them, from Davy's grey to Fuzzy Wuzzy — and kept separate rather than folded into a base. This is the finer reading: it distinguishes tones the twelve categories flatten together, at the cost of a long tail where most names sit on only a handful of objects. The twelve categories are these tones, grouped.
Two different times run through a collection. There is when each object was made, and there is when the museum took it in. They are rarely the same, and the distance between them is collecting policy made visible.
Above the line: when the objects were made. Nearly every published object carries a production date, but only a quarter carry an exact year — the rest are spans, some a century wide. Rather than pinning those to their first year, each object spreads a single unit of weight evenly across the years it might have been made in. An object dated 1937 puts its whole weight on one year; one dated 1900/2000 puts a hundredth on each of a hundred. So height is share, not certainty, and vague dating shows as flatness rather than as a false spike. The darker band is the objects dated to a single year, which says as much about when record-keeping became precise as about the objects themselves.
Below the line: when they entered the collection. An acquisition date is a point, not a span — the museum took the object on one day — so this is a plain count with no weighting. It is deliberately spiky: collections grow in events, and a single bequest or purchase can bring in hundreds of objects on one date. Note that the line begins only in 1904, a year after the museum was founded, and that around a quarter of published objects carry no acquisition date at all.
Reading the gap. Where the upper shape runs ahead of the lower one, the museum was collecting objects made long before — the 1980s took in more objects than any other decade, most of them decades old already. After 2010 the relationship inverts. And both shapes thin sharply through the 1930s and 1940s: the collection neither acquired much then nor holds much made then.
Each bar is a decade of acquiring. It is divided by how old the objects were when the museum took them in, and its height is the number of objects — so the decades that collected little are visibly slight, rather than given the same authority as the ones that collected thousands.
The age is a range, not a number. An object recorded as made between 1900 and 1950 and acquired in 1907 was somewhere between nought and seven years old — the 1950 was an open-ended note, and the purchase itself is evidence. Each object is banded on the midpoint of its possible range, which is an estimate. The purple rule under a bar is the share whose dating is too loose to band confidently, and it is widest exactly where the collection is largest.
It changes direction more than once. The museum opened buying antiques, turned to the design of its own moment within twenty years, swung back to the historical, and reached its most contemporary in the 1990s — when half of everything acquired had been made that same year. Since 2010 it has been turning again.
The other two indexes sort the collection into categories. This one does not sort it at all. It measures how close any two photographs sit in a space built by a machine that has looked at several hundred million captioned images — and lets you walk from one object to its nearest neighbours, and on from there.
There is no overview here, and that is the point. Colour and time can be summarised in a single band, because every object has a colour and a date. Resemblance is relational: it only exists between things. So there is no picture of the whole — only a starting object and what sits nearest to it, and then whatever sits nearest to that.
What the machine is doing. Each photograph is turned into a list of 768 numbers by CLIP, a model trained on images paired with the text found near them online. Images described in similar ways end up close together. Nothing here was catalogued: the model was never shown this collection, never told what a decorative tile is, and has no access to any record. It compares pictures.
What it therefore cannot know. It brings the internet's associations with it, including its blind spots and its clichés, and it will happily group objects for reasons that have nothing to do with design history — a shared backdrop, a similar crop, a photographic convention. It is not a judgement about the objects. It is a measurement of their pictures, and it is worth disagreeing with.
A photograph is mostly not the object. Studio backdrops, shadow and floor can take up more of the frame than the thing being photographed, and a colour reading from the whole image describes the photography as much as the collection. The background is removed before the pixels are clustered.
