Build your own HackerTube

1. Get the export

The official data lives at hackertube.org/about ("Download the data"). It is a 17.8 MB zip, hackertube-data.zip, released under ODC-By 1.0: use it for anything, credit HackerTube and the open-index/hacker-news dataset.

File	One row is	Key
videos.jsonl	a video (6,454 rows, best-scored first): id, title, channel, summary, why HN likes it, format, level, shelf life, topics, categories, pull quote, HN stats, score	vid
mentions.jsonl	a labeled HN comment (39,834 rows): comment id and link, author, month, story title, stance, why, quote, cleaned text	hn_comment_id + vid
taxonomy.json	16 areas and 150 categories with synonym-rich descriptions	category id
related.jsonl	"the same HN people also recommend", for 3,778 videos (counts only, no usernames)	vid
embeddings.npy + embedding_ids.json	a 384-dimensional bge-small vector per video, L2-normalized, so semantic search and "similar videos" come free	row index
DATA.md	documentation written for AI agents: joins, gotchas, example records, licensing

2. Hand it to an agent

The export is designed to be given to Claude Code or a similar agent with one sentence of intent. Some prompts that work:

• "Build a static site with one page per video and per topic, using the HN comments as the page body." (This site.)
• "Email me, every Monday, three videos from topics I list that I have not seen, with the best HN quote for each."
• "Make a terminal browser: fuzzy search over titles and summaries, open in mpv."
• "Which channels have the most strong recommendations per video? Which years produced the most evergreen talks?"
• "Re-rank everything for me using my own liked list and the embeddings."

3. The gotchas (from DATA.md, confirmed while building this)

• Only stance ∈ {rec, strong_rec} are recommendations. About 25% of videos have hn.recommenders == 0; decide whether to show them.
• At most 20 mentions per video are included, so hn.mentions is capped at 20.
• shelf_life == "news" marks videos that only mattered around one event; most interfaces should hide them by default (407 videos).
• Some talks were uploaded twice under different ids with near-identical titles (18 duplicate titles). Dedupe by normalized title if it matters to you.
• Thumbnails need no API key: https://i.ytimg.com/vi/<vid>/hq2.jpg is the mid-video frame HackerTube uses; hqdefault.jpg is the uploader's custom one. Embed with https://www.youtube-nocookie.com/embed/<vid>.

4. YouTube metadata, legally

Durations, view counts, publish dates and whether a video has been deleted are not in the export, because YouTube's terms do not allow redistributing them. Every record has the video id, so you fetch them yourself with a free YouTube Data API key: videos.list takes 50 ids per request, all 6,454 videos cost about 130 units of a 10,000-unit daily quota. About 6% of videos turn out to be gone. Keep the key private, show the data with attribution to YouTube, and refresh it from time to time.

5. What this edition adds, if you want to copy it

• A static site generator (Astro) that renders every video, topic, area, channel and format page at build time, so search engines and AI crawlers can read them. The original is a single-page app with no sitemap.
• Per-video "recommended in which years" computed from the mentions, shown as a strip on every card.
• Top-8 similar videos precomputed from the embeddings with one matrix multiply, folded into each video page, and reused in the browser for a no-account "because you liked" row.
• Full-text search over titles, summaries, tags and comment text with Pagefind, built from the HTML itself.
• A refresh pipeline that queries the Hacker News search API for new YouTube links after the export date and applies the same three-people, three-months rule.

Credits and licensing for this site are on the data page.