Art Cal (Making × Witnessing)

Built by Hannah Levinson • more at hrlevinson.com

A personal calendar for making and seeing art in New York.

Live: https://nyc-art-cal.vercel.app

Subscribe from Google Calendar or Apple Calendar:

• All events → /feed.ics
• Just shows to attend → /feed-attend.ics
• Just classes and workshops → /feed-practice.ics

---

Why I built it

I wanted a personalized instrument for moving through the artworld and my own art practice in New York.

What it is

• A weekly calendar of events across venues including theatre, dance, film, music, workshops, readings, and more.
• Two modes: Practice (classes and workshops, mode: make) and Attend (shows and screenings, mode: witness).
• A picks / shortlist system with passphrase-based cross-device sync — star on my phone, see it on my laptop.
• A self-updating ICS feed so the whole thing lives in Google Calendar alongside everything else.

What it does

• Groups events by week, one week at a time by default. Auto-focuses on the current week; arrows to move; jump-to-week dropdown.
• Every event carries structured metadata: category, mode, start / end time, category-color left border.
• Weekly summary strip: event count, cost range, making %, picked count and cost.
• Past events dim and hide by default with a "show past" toggle.
• Places to make / see things: every venue as an ongoing resource, collapsed below the calendar so it doesn't wall off the events.
• Cross-device picks sync via Supabase, keyed to a passphrase hash — no account, no email.
• Filterable exportable feeds that GCal auto-refreshes.

The categories

• sound — concerts, sound art, opera
• dance — dance shows and classes
• film — screenings, workshops
• tech — live-coding, generative visuals, AI, hardware
• make — printmaking, book arts, woodworking, darkroom
• stage — theatre, performance
• word — writing groups, readings
• circle — social practice, communal meals, meditation

Each shows in a distinct color so I can scan a week in a second.

What it is not

• Not a listings site. It's a curated personal map. Every entry landed here because I directly verified it on the venue's own site — not from a scraped aggregator.
• Not a social product. The picks-sync is one person's picks shared across their own devices. No feed, no other people, no sharing links.
• Not automated. New events go in by editing JSON in the GitHub web editor. No scraping cron, no email-parsing pipeline. The friction of adding an event forces me to consider whether it belongs.

How it's specific to me

The engine is generic. What's on the calendar reflects my personal practice and preferences.

Fork the code, rewrite the JSON.

How it's built

• Vite + React + TypeScript, strict mode, no CSS framework — hand-written stylesheet on CSS custom properties.
• All events in JSON — src/data/events.json, spaces.json, fall.json. I edit these directly in GitHub's web editor and commit — no admin UI, no forms.
• Supabase for picks sync. SHA-256 hash of a user-chosen passphrase is the row key. The passphrase is never sent to the server, only its hash. Nothing else is stored in the backend.
• Build-time ICS generation. A tsx script (scripts/generate-feeds.ts) runs after Vite build and emits three .ics files into dist/, served with Content-Type: text/calendar from Vercel via vercel.json.
• Stable event UIDs derived from date + slug of event name, so calendar apps update existing events instead of duplicating them on each refresh.
• Today-aware everywhere. Days-until countdowns on events, current-week highlight, past-event dimming — all keyed to today() at load.
• Deployed on Vercel with continuous deploy from main. Editing JSON on GitHub triggers a Vercel rebuild within a minute; GCal picks up feed changes on its next refresh (typically 12–24 hours).

Scanning

A scheduled GitHub Action (scripts/scan-events.ts) sweeps the venue list in scripts/scanner/venues.ts and extracts upcoming events. It tries the cheapest, most reliable source first and only falls back when it must:

1. Published iCal feeds (icsUrl on a venue) — parsed directly, no LLM.
2. JSON-LD structured data on the listing page.
3. Detail-page crawl — follows event links to pages that carry JSON-LD even when the listing doesn't.
4. Platform parsers — server-rendered Squarespace event lists and WordPress "The Events Calendar" markup.
5. LLM extraction — last resort for JS-rendered venues with no structured data.

Every candidate passes gates (real future date present in source, title tokens present, valid category/mode) and a fuzzy dedupe check; dense performance runs collapse to one entry.

No-pay LLM setup

Step 5 uses whichever free API key is present, preferring the largest free tier. Set one as a GitHub Actions secret:

• GROQ_API_KEY — Groq free tier (thousands of requests/day; covers every LLM-needing venue each run). Recommended.
• CEREBRAS_API_KEY (+ optional CEREBRAS_MODEL, default llama-3.3-70b) — Cerebras free tier; a strong alternative if Groq signup is down.
• GOOGLE_API_KEY — Gemini free tier (~20 requests/day; venues rotate across runs via scan-state.json).
• OPENAI_COMPAT_BASE_URL + OPENAI_COMPAT_API_KEY (+ optional OPENAI_COMPAT_MODEL) — any OpenAI-compatible endpoint (OpenRouter, Together, …).

With no key set, steps 1–4 still run, so feed/JSON-LD/CMS venues keep updating for free.

Editing

Any event or venue lives in src/data/. Edit in the GitHub web editor, click Commit — Vercel rebuilds automatically and your subscribed calendar reflects the change on its next refresh cycle.

Where it lives

• App: https://nyc-art-cal.vercel.app
• Author: hrlevinson.com