📊 See open-source project rankings

📱 Get the AI Rank app

HackerTube is a public catalog at hackertube.org of YouTube videos that keep showing up in Hacker News comments. It is not a YouTube clone that hosts video files. Playback goes through YouTube's privacy-enhanced embed (youtube-nocookie.com). The site's own copy describes itself as "The YouTube videos that Hacker News keeps telling each other to watch."

On November 2, 2026 (America/Los_Angeles), the live /api/stats endpoint reported 6,454 videos, 39,834 labeled HN comments, 15,041 recommenders, 150 topic categories in 16 areas, covering HN years 2007–2026. Those counts match Andrej Karpathy's October 31, 2026 write-up and the on-site data export dated the same day. For the launch story (Claude Code / Opus 5.5 weekend build), see our news note: Karpathy launches HackerTube. This guide focuses on what the site does, how to browse it, and how to reuse the dataset.

What HackerTube is for

YouTube's home feed optimizes for the next click. HackerTube's About page states a different goal: surface talks and documentaries that HN commenters keep pointing to over years, not one viral spike. Ranking on the site uses how many separate months people recommended a video. A link that returns across years outranks a single burst.

The pipeline, as the About page describes it:

  1. Scan a public full HN archive for YouTube links (the data page credits open-index/hacker-news on Hugging Face).
  2. Keep a video only if at least three different people linked it with real commentary in at least three different months. One-off drive-bys drop out. About text says that leaves roughly 7,000 candidates before final packaging; the live catalog we hit is 6,454.
  3. Label each comment with an AI model for stance (recommend, cite, joke, criticize), plus a short quote and topics.
  4. Write a card per video (summary, why HN likes it, type, shelf life) and sort into a shared topic list.
  5. Rank by lasting recommendations across months.

Karpathy's post adds process detail: Claude Opus 5.5 in Claude Code coordinated the work; Sonnet ran over comments in a Workflow; spam was filtered down to about 40K comments that recommend a YouTube video (close to the 39,834 mentions in the export). He says the only manual steps were creating a YouTube Data API key, booting a VPS, and buying the domain (and, in an HN reply, setting DNS).

How to browse the site

No accounts. Likes, Watch later, and History live in browser localStorage only. Videos play in a privacy-enhanced YouTube iframe.

Useful surfaces we verified in the shipped UI:

  • Home. Mixes HN classics, lower-view "hidden gems," recent favorites, unexplored topics, and (after you like a few videos) similar items plus "HN people also recommend" picks. Hover a card to see why it is in the feed.
  • Search. Topic-aware search, not only keyword match. The About page's example is searching "distributed systems" and still finding talks that never use those words.
  • Categories. Sixteen areas with 150 categories. Live /api/browse examples include Software Engineering & Languages (1,604), AI, Machine Learning & Data (718), Systems, Networks & Security (956), and Arts, Comedy, Music & Film (1,731). A video can appear under more than one topic, so area counts are not meant to sum to 6,454.
  • Longs. Nav link filters to videos over an hour (/browse?dur=xlong).
  • Playlists. Curated lists (Claude-authored intros). /api/playlists returned 27 playlists at access time, including starters such as "The HN canon" and "Talks every programmer should watch."
  • Video pages. Show the retained HN comments with links back to threads.
  • Keyboard shortcuts. Press ? for the full map; / focuses search; j / k move through videos.
  • Mid-video frames. Settings default to showing a frame from the middle of the video (YouTube hq2.jpg or a local /frames/ copy) instead of the channel's uploaded thumbnail. Karpathy calls this the "50% thumbnail."

The footer disclaimer is clear: HackerTube is independent and not affiliated with YouTube, Google, Hacker News, or Y Combinator.

Dataset export and self-hosting your own view

Open hackertube.org/data or About → download. The export metadata (/api/export, generated 2026-10-31) lists a zip named hackertube-data.zip (17,842,799 bytes). We confirmed https://hackertube.org/download/hackertube-data.zip returns HTTP 200 with that content length.

Files inside (from the export API and About table):

File What it holds (site wording)
videos.jsonl 6,454 rows: YouTube id/link, title, channel, summary, why HN likes it, type, level, shelf life, topics, categories, pull quote, HN stats, quality score
mentions.jsonl 39,834 labeled comments: ids, links, author, month, story title, stance, quote, cleaned text
taxonomy.json 16 areas / 150 categories with synonym-rich descriptions
related.jsonl "HN people also recommend" edges for 3,778 videos (counts only, no usernames)
embeddings.npy + embedding_ids.json bge-small vectors, 384 dimensions, 6,454 ids
DATA.md Agent-oriented docs: joins, gotchas, stance definitions, example code, licensing

License on the data page: ODC-By 1.0, same family as the HN archive it derives from. Credit HackerTube (and the open-index dataset). Comment text remains the authors' work. Labels and cards were written by Claude, so some will be wrong.

Not in the zip: YouTube descriptions, view counts, durations, or publish dates. YouTube's API terms block redistributing that metadata. Every row still has the video id; the data page says a free YouTube Data API key can fetch the rest in about ceil(6454/50) requests, with steps in DATA.md.

Suggested path from the site itself: unzip, point Claude Code (or similar) at the folder, and ask it to read DATA.md before building a daily mix UI, a weekly email, a terminal browser, or an analysis. The Data page even ships a copyable agent prompt for a small local web app with topic/length filters and watch history.

Karpathy did not publish the live server application as an open repo in the sources we checked. Reuse starts from the data export, not a drop-in clone of hackertube.org's backend.

HackerTube vs YouTube recommendations

YouTube home / related HackerTube
Signal Watch-time and engagement models owned by YouTube Recurring HN comment links, labeled and ranked by months of recommendations
Accounts Usually signed-in personalization None; state stays in localStorage
Thumbnails Channel-uploaded art (often clickbait) Optional mid-video frame by default
Scope Entire YouTube corpus ~6.5k videos that cleared the multi-person, multi-month bar
Forkability Closed ranking stack ODC-By dataset + embeddings for your own agent

Neither site replaces the other. HackerTube still depends on YouTube for playback and metadata refresh. It is a discovery front end for a narrow, community-filtered slice. If your taste matches long-running HN recommendations, the Home feed and Categories are a practical starting point. If you need trending Shorts or mainstream entertainment, stay on YouTube.

FAQ

Is HackerTube free?

Yes. The website needs no login. The dataset zip is free under ODC-By 1.0 with attribution.

Does it host the videos?

No. It catalogs and embeds them. Files stay on YouTube.

How many videos are on HackerTube?

6,454 on the /api/stats and /api/export responses we fetched on November 2, 2026. Re-check the live API if you need a later figure.

Can I build my own HackerTube?

You can reuse the published zip and DATA.md with your own agent or scripts. There is no official public dump of the production server code in the sources cited below.

Sources