Look for an open dataset of real user interface (UI) screenshots, of the kind Rico or a component library once offered, and the search comes back almost empty. It is not thin in the way a young category is thin, where a few small projects promise more later. It is empty in the way a category looks when the obvious thing has not been built, and the reasons become visible once you look at what exists instead.
The strongest signal is the best result. Among open UI pattern and screenshot datasets, waltteri/desktop-ui-dataset is the closest thing to what the search was looking for: an annotated desktop UI screenshot dataset in PASCAL Visual Object Classes (VOC) Extensible Markup Language (XML) format, 8 stars, no LICENSE file, last updated 2020-12-27. It is nearly six years stale, small, and unlicensed. That is the best hit, not a middling one. Everything else in the category is a live tool or a curated reference site, not a corpus a team could download, version, and build on.
First principles
Four facts explain the gap before any project is named.
- A dataset is a fixed copy. It is a versioned snapshot that others download, cite, diff, and reuse. Reuse at scale needs a license that grants it.
- A screenshot of a commercial product reproduces someone else’s work. The person who captures it did not make the design in it, so they cannot grant reuse rights over it on their own.
- A live tool stores nothing. It fetches a page when asked and discards it, so it never redistributes anything. For the same reason it leaves nothing fixed to diff, train on, or cite.
- A measurement is not a copy. A hexadecimal color value, a type scale, or a spacing step describes a design without reproducing its pixels.
What follows: an open corpus of real UI has to give up either the images or the freedom to redistribute them. The rest of this essay shows the market choosing the second, and argues for the first.
What the search actually found
Three projects sit in the open UI dataset category, and each fails the “durable, redistributable corpus” test in a different way.
waltteri/desktop-ui-dataset is the only literal screenshot dataset in the set, and it stopped moving in December 2020.
inbn/component-gallery, “The Component Gallery,” is closer to a reference catalog than a bulk dataset. It is a curated, hand-maintained collection of real components pulled from best-in-class design systems, backed by Astro and Airtable, with 348 stars, no LICENSE file, and a last update on 2026-06-12. It is actively maintained and useful to browse. It is still editorial curation, not a training-scale corpus, and it ships without a license that would make reuse explicit.
INSANE0777/Awwards-mcp is the most recent and most telling entry: 47 stars, MIT-licensed, last updated 2026-09-20. It is a Model Context Protocol (MCP) server, not a static dataset. It gives coding agents design inspiration from award-winning sites through live screenshot search, “design DNA” extraction, and page capture. It is pitched as a Mobbin alternative with no application programming interface (API) key required. It is the closest thing the research found to an open Mobbin or Refero substitute, and it is a live scraper wired up as an agent tool, not a downloadable corpus.
Read together, these three projects say the same thing three ways: nobody has shipped a large, current, freely redistributable set of real UI screenshots or patterns. What exists is years stale, a small curated catalog without a clear license, or a tool that fetches fresh material on demand instead of storing it.
The likely reason is copyright and ToS risk
None of these projects states why it stops short of a redistributable corpus, so what follows is the likely explanation, not a proven one. The shape of the gap points the same way every time. Dataset projects with an explicit LICENSE file are rare in this category. Static image collections are rare. The entries that exist and stay active are the ones that never store the images at all.
That is the pattern copyright and terms of service (ToS) risk would produce. Screenshots of commercial products are not neutral assets. Redistributing them at scale, bundled into a downloadable dataset with a permissive license, creates a different legal exposure than fetching a page live for a single inspection and discarding it. A live scraper can plausibly argue that each fetch is closer to normal browsing. A static, redistributed archive of thousands of screenshots from named commercial products looks more like the thing copyright and platform terms of service exist to prevent. waltteri/desktop-ui-dataset and inbn/component-gallery both ship without a LICENSE file, which is consistent with maintainers who will not make an explicit reuse claim over material they did not originate.
Live scrapers fill the gap, not corpora
INSANE0777/Awwards-mcp deserves a closer look because it is the shape the market has converged on instead of a dataset. It fetches screenshots and extracts a “design DNA” summary at query time, per agent request. Nothing is stored centrally, nothing is versioned as a release, and nothing is redistributed to a third party as a downloadable file. The structure sidesteps the redistribution question by never redistributing: each user’s agent makes its own live request against the live site.
This is a real and useful pattern for an individual agent doing individual research. It is not a substitute for a corpus. A live scraper cannot be diffed against a prior version to see what changed. It cannot be used to train or evaluate a model offline. It cannot be cited with a fixed snapshot the way a dataset release can. And every user of the tool re-scrapes the same sites. That is worse for the sites being scraped than a single, well-governed archive would be, and worse for reproducibility than a pinned dataset release. The live-scraper pattern is a workaround for the absence of a dataset, not a replacement for one.
Galleries are not datasets either
Two adjacent products show the same gap from the gallery side. Both show why “a lot of screenshots in one place” is not the same as an open corpus.
CollectUI is a long-running Dribbble and Daily UI screenshot gallery with no structured output: no tokens, no DESIGN.md, no API, and no Really Simple Syndication (RSS) feed. It has no /about page; that address returns a 404 (not found) error. The individual design pages carry no visible post dates and no visible attribution back to the Dribbble shot owners. Traffic figures cited for the site (2,555 unique daily visitors and 11,243 pageviews in August 2026), and a claim that its parent network, Panda, routes over 1.2M pageviews a month to Dribbble, are unverified; the research could not confirm them independently. The site is funded through sponsorship and backlink deals rather than user pricing. Whatever CollectUI is, it is not a corpus a team could adopt. There is nothing to export, no license terms attached to the images, and no attribution trail for the designers whose work fills the pages.
Refero’s public gallery, styles.refero.design, is the opposite failure mode. It is well built and well attributed at the level of individual brand pages, but it is not open. The gallery describes itself as offering “2,000+ AI-readable design systems from leading product websites,” drawn from a larger paid platform with 142,000+ web and iOS screens and 12,000+ user flows. Refero’s pricing page describes the free tier as access to roughly 3% of the full collection. The same page lists the paid tier at $10 a month billed annually, or $12 per seat per month billed annually for teams; another Refero page showed a different price for what appeared to be the same plan.
Whatever sits behind that paywall is not redistributable to a third party under any license terms found in the research. It is a product to browse and query, not a corpus to fork. Refero also discloses that its per-brand pages are interpreted and normalized rather than raw extraction. That is a separate fidelity question from the licensing one, but it reinforces the same point: even the richest artifact in this research is a proprietary product built on other companies’ UI, not an open dataset anyone else can build on.
A dataset that could be legally shipped
The proposal that follows is not to build a bigger screenshot archive and hope the risk resolves itself. It is to stop shipping images at all.
A derived-tokens-only corpus removes the redistribution problem at its source. Instead of storing and distributing screenshots of a product, the corpus would store only what has been extracted and abstracted from it: colors, typography scales, spacing values, and component descriptions, written as DESIGN.md and Design Tokens Community Group (DTCG) token files rather than pixels. A token file derived from a public page is a different kind of artifact from a copy of that page’s rendered output. It is closer to a design critique or a technical description than a reproduction.
Three conditions would need to hold for this to be defensible in a way a raw screenshot archive is not.
- Attribution. Every entry names its source brand and source Uniform Resource Locator (URL), the way a citation names a source instead of reproducing the whole work.
- Provenance. Every entry states how it was produced, the same discipline argued for in Nobody publishes extraction accuracy. A reader then knows whether a token was scraped from the page’s Document Object Model (DOM), written by hand, or inferred by a large language model (LLM), and can weight it accordingly.
- Brand opt-in and opt-out. A named brand can ask to be added, request corrections, or ask to be removed entirely, and that request is honored.
The last condition is the one none of the sources in this research offer. CollectUI gives no attribution and no removal path. Refero’s gallery is built from commercial products without an opt-in or opt-out mechanism found in the research. A dataset that gives brand owners a real lever is a different kind of artifact from one that treats every public page as fair game by default.
None of this produces a screenshot dataset, and it should not be sold as one. It produces something narrower and more defensible: a corpus of what a site’s visible design choices reduce to in token form, attributed to its source, tagged with how it was derived, and revocable by the brand it describes.
What Velofy is doing next
Velofy intends to explore building this derived-tokens-only corpus rather than a screenshot archive, starting from public sites where attribution and provenance can be recorded cleanly. We intend to design an opt-in and opt-out path for brand owners before any entries are published at scale, not after. Until that groundwork is in place, Velofy will keep treating the open UI dataset gap as unresolved rather than implying that any live-scraper tool or paywalled gallery already fills it.