You open YouTube for one good lecture and end up choosing between shouty thumbnails and topics you’re already sick of. HackerTube offers a different way in: videos that Hacker News members have kept coming back to for years, with the context of their recommendations. Andrej Karpathy built the service with Claude in about two hours, by his own estimate. You can already try the catalogue, and the way it’s put together hints at how to build useful products on top of existing platforms.
HackerTube doesn’t replace the video platform. It changes how you choose what to watch: YouTube stores and plays the videos, and a separate catalogue helps you decide what to spend the evening on. For developers and entrepreneurs it’s a useful example: sometimes a new product starts not with its own content but with a different selection rule.
HackerTube: a video catalogue instead of an endless feed
According to the project page as of 4 November 2026, the catalogue holds 6,454 videos, 39,834 labelled Hacker News comments and 150 topics across 16 broad areas. It’s built on discussions from 2007–2026.
According to the project’s description, the home page mixes several kinds of selection: community classics, little-known finds, recent recommendations and topics the user hasn’t explored yet. Once you’ve liked a few videos, similar ones appear, along with videos recommended by the same Hacker News members. On a computer you can see why a card appeared by hovering over it; we didn’t check how to view the explanations on a phone.
So recommendations haven’t gone anywhere. What’s changed are the input signals and the rules for what gets shown. Rather than promising “YouTube without algorithms”, it’s more accurate to talk about a catalogue with its own recommendation system.
There are filters by length and type of video, topic categories, playlists and semantic search. On the public home page we saw, for example, Bret Victor’s talks on the future of programming, Rich Hickey’s talks and the Statistics for Hackers lecture. That gives a sense of the selection’s leaning: it’s a good place to look for technical talks, rather than a replacement for everything YouTube offers as entertainment.
A link in a comment isn’t necessarily a recommendation
The most substantial part of HackerTube is the preliminary data processing. A plain list of YouTube links wouldn’t have been enough: people share videos to praise them, argue, joke or show a bad example.
Judging by the project page, the comment archive was first checked for YouTube links. Then they kept the videos that at least three different people had mentioned with a substantive comment, across at least three different months. A filter like this should weed out random links and one-off spikes of attention.
Next, Claude analysed each comment together with the discussion’s title and identified the attitude towards the video: a recommendation, a reference to a source, a joke or criticism. On that basis came short descriptions, topic labels and explanations of what the video is valued for. Ranking uses, among other things, the number of different months in which it was recommended.
You can compare it to a bookshelf put together by specialists you know. What’s interesting isn’t a book someone mentioned once, but one that different people keep recommending for years. The algorithm makes that kind of signal searchable.
That said, the shelf is still someone else’s. Hacker News tastes aren’t the tastes of the whole internet, repeated recommendations don’t prove the content is reliable, and the model’s labelling can be wrong. Be especially careful with material on health, finance and other subjects where popularity is no substitute for fact-checking.
What “built in two hours” actually means
Karpathy puts the time to build the project with Opus 5.5 at about two hours. The HackerTube page credits Claude Opus with the work on data processing, search, recommendations and the site, and teams of Sonnet agents with the comment labelling, the video cards and the topic list.
That’s a description of one specific run, not a promise that anyone will get a result in two hours. You can’t automatically swap in your own task, data volume and reliability requirements. The primary sources we studied contain no reproducible report that would let anyone independently check the whole process and its cost.
The strong point of the case, in our view, is a well-chosen product boundary. There was no need to build video storage, an upload system or delivery infrastructure of its own. The effort went into what the user was missing: selection, context and search.
If you want to repeat the approach, it’s better to start by framing the task rather than with a “make a clone” command. For a first look at the tool, our guide to Claude Code will come in handy, and for organising the work, 40 rules of vibe coding. Generating an interface quickly doesn’t remove the need to check what ended up inside it.
No account doesn’t mean no YouTube and its limits
According to HackerTube, your history, likes and “watch later” list are stored only in the current browser. No sign-up is needed. That’s handy for a quick start, but don’t count on automatic syncing between devices.
Videos play through the embedded YouTube player in privacy-enhanced mode. According to YouTube Help, views in this mode don’t affect YouTube’s own personalisation. But that’s no promise of complete anonymity and no guarantee of no ads: the help page explicitly allows for non-personalised ads.
There’s one more technical boundary. The rel=0 parameter in the embedded player doesn’t switch off related videos completely; it limits them to the same channel. So the catalogue’s own home page and the recommendations inside the player are different things.
For anyone building a similar product, the platform’s rules matter. According to YouTube’s documentation, current as of 4 November 2026, applications must not block ads or the player’s standard features; a service that repeats the YouTube user experience must offer value of its own and must not mislead people about where it comes from. This isn’t a legal assessment of HackerTube, but it is a reason to check the requirements separately before launching a catalogue of your own.
You can build on the data, but it’s not a ready-made clone
HackerTube offers its prepared data for download: video records, labelled comments, the topic structure, links between videos and vectors for semantic search. The bundle includes DATA.md, a description of the formats and quirks for working with an AI agent.
The project states an ODC-By 1.0 licence with a requirement to credit HackerTube. The licence itself governs rights to the database, not automatically to every item included in it. Separately, the project says YouTube metadata such as descriptions, view counts and duration isn’t included in the export; you’re expected to fetch the information you need yourself using the video IDs. Before using it, read the terms of the export itself and YouTube’s requirements, rather than treating the archive’s availability as permission to do anything you like.
From this data you could ask an agent to put together a themed library of talks, a personal list of lectures on distributed systems or a selection for a study group. These are possible uses of the dataset, not features we tested in the finished service.
For an IT community, a similar approach could work with members’ public recommendations: keeping not just the link but an explanation of who the material helped. Private conversations shouldn’t be used for an experiment like this without the participants’ permission.
Start with one topic, not your own YouTube
We looked at the home page and the project description; we didn’t test playback, personal recommendations or the data download.
To judge the service for yourself, open HackerTube, pick a topic you know and compare a few cards with what your feed usually offers. Look not only at the video but also at the comments the recommendation is based on. That makes it easier to see whether the community’s taste matches yours.
And if you’re interested in a project of your own, give the agent a narrow task: “Build a catalogue of publicly recommended talks on my topic. Show the reason for each recommendation and a link to the discussion. Don’t download the videos; for playback, use supported embedding in line with the platform’s rules.”
The test is simple: if even one of those cards led you to material you’ve long wanted to find, the catalogue makes sense for you in particular.
