Art Daily

Art Daily picks a real artwork and writes a readable article about it, grounded in the Met Museum and Wikipedia so the dates, attributions and techniques come from records instead of the model's imagination. It reads in Spanish or English, and you can translate any article on demand. Two tabs hold the app: Articles, with an editorial reader and a chatbot tied to each artwork, and Exhibitions, a city-based explorer of current shows. Everything runs on a home server, so there are no AI bills and nothing leaves the house.

Highlights

Grounded articles, not invented ones. Every piece is anchored to museum and Wikipedia data before the model writes anything, so the facts trace back to a source. When the sources don't cover something, the text says so rather than filling the gap.
Search your library, or create something new. One box filters the articles you've already made; type a name it hasn't seen and it offers to generate that one. The same field does both jobs without making you pick a mode first.
Search by artist, not just by title. Type "Monet", "Picasso" or "Basquiat" and you get that artist's actual works. Historical artists pull images from the Met and Wikimedia Commons; modern artists under copyright (Basquiat has 72 catalogued works and zero free images) fall back to a web image search, so they still show up.
Bilingual, with translation on demand. The interface switches between Spanish and English, and each article keeps a separate version per language. Open one written in the other language and a single click translates it, grounded the same way the original was.
Exhibition finder. Search a city and get a de-duplicated list of current shows, with finished ones dropped by date and a detail view for each.
A chatbot per artwork. Ask about the piece, request more by the same artist, or highlight a passage in the article to ask about that exact sentence. Answers cite their sources, and you can pin the good ones back onto the article.

How it works

The front end is a no-build site: plain HTML, CSS and JavaScript with an editorial serif look, talking straight to a set of n8n workflows. No framework, no bundler, one script file.

The workflows do the research. They pull facts from the Met Museum API and Wikidata/Wikipedia, and when those come up empty they fall back to a self-hosted SearXNG instance for open-web answers. The writing goes to a local LLM (qwen2.5 on LM Studio), with the model configurable per task. The rule is the same everywhere: answer from verified sources, and admit it when the answer isn't there.

Generation runs in two phases so it feels quick. The first finds and normalizes the artwork (image plus a fact sheet) and shows it almost immediately; the second writes the full article in the background and drops it in when ready. The daily pick, the query generator and the chatbot all sit on this pipeline.

The artist search is the part I'm most pleased with. Working out whether you typed an artist or a title is half the problem: it checks whether the name resolves to a person with an art occupation in Wikidata, and when that's ambiguous (the surname "Velázquez" returns a metro station and a street before the painter) it asks the Met, which attributes works to artists and links them back to Wikidata. Once it knows the artist, it queries Wikidata for everything that artist created, prefers works that have a free image, and reaches for SearXNG only when there isn't one. Images cascade the same way throughout: Wikimedia Commons first, then a Wikipedia thumbnail, then web search as a last resort.

Articles and exhibition history live in the browser for now, capped at the most recent 50, so there's no database to babysit yet.

Screens and demos

Home grid

Generated artwork articles feed

The editorial reader

Article reader

Search by artist

Searching by artist

Chat with starter questions

Per-artwork chat

Select to ask

Exhibitions searcher

Bilingual and translate on demand

Spanish and English with on-demand translation

Tech stack

Vanilla HTML/CSS/JS (no framework, no build) · n8n · local LLM (qwen2.5 on LM Studio, swappable per task) · SearXNG · Met Museum API · Wikidata SPARQL and Wikipedia · light and dark themes · Spanish and English.

Status and what's next

Live, fully self-hosted, and bilingual. Article and exhibition history sit in the browser today. Next: moving content into a database, and a proper "artwork of the day" feed.