HackerTube: Karpathy Built a YouTube Alternative in ~2 Hours
Karpathy built HackerTube in ~2 hours with Claude Code: 6,454 YouTube videos picked by HN commenters, anti-clickbait thumbnails and account-less ForYou picks.

Andrej Karpathy — former Tesla AI director and OpenAI founding member — spent a Friday night building HackerTube, a YouTube look-alike with an unusual source of taste. Its catalog is 6,454 videos that Hacker News commenters actually recommended to each other, harvested from roughly 39,834 comments posted between 2007 and 2026. No algorithmic feed, no watch-history profiling — just two decades of "watch this" from people who argue about compilers for fun.
He announced it on October 31, 2026, with a short writeup on his blog, and the Hacker News thread exploded: 2,877 points and 637 comments within days. The site got most of the attention, but the build log is what developers kept quoting. Karpathy says he wrote none of the code in any traditional sense — he talked to Claude, and Claude did everything else.
Below: what HackerTube is, what it feels like to use, how it was built in about two hours, and why a family of copycats appeared within days.
Why Karpathy built it: an escape from the slop
In the writeup, Karpathy describes opening YouTube and finding "an enshittified wall of slop" — recommendations chasing watch time, clickbait thumbnails, an endless middle of mediocre filler. His complaint is not that YouTube lacks good content. It is that the surface you actually see is tuned for engagement, not quality.
His fix sidesteps the algorithm entirely. Instead of a recommender trained on viewing behavior, he used one assembled from text: the explicit video recommendations scattered through Hacker News comments over nearly two decades. The taste being aggregated is unusually specific — HN commenters recommending videos to other HN commenters — and that specificity is the point.
Why not just use YouTube's own recommendations?
Because the problem is the objective, not the data. YouTube's recommender optimizes for watch time and clicks because that is what the platform sells; a pile of "this is genuinely worth your evening" comments optimizes for something closer to taste. Karpathy swapped the objective instead of tuning the old one.
It also explains why nobody mistakes HackerTube for a business. There is no engagement loop to optimize, because engagement was never the goal.
Inside HackerTube: what you actually get
The site plays it straight: a YouTube-style grid with search, topic navigation and a player. The interesting decisions are in the details.
| Aspect | How HackerTube handles it |
|---|---|
| Catalog | 6,454 videos across 150 topics, uploaded between 2007 and 2026 |
| Thumbnails | Frame from the 50% mark of the video, not the uploader's |
| Recommendations | Account-less "ForYou", state kept in the local browser session |
| Accounts | None — there is no login anywhere |
| Cost | Free |
The thumbnail rule is the standout. The image you see before clicking is a mid-video frame that nobody designed to trick you, which fixes clickbait at the catalog level instead of asking you to judge each cover yourself.
Pair that with keyboard shortcuts and an account-less ForYou feed, and the site feels closer to a well-kept archive with a friendly face than to a platform fighting for your whole evening. The content's shape is predictable too: talks, documentaries and internet history dominate — HN loves what HN loves, with the occasional left-field pick.
How Karpathy built it in about two hours
This is the part most readers came for, because the build story is unusual even by 2026 standards. Karpathy calls it a Friday night project — roughly two hours end to end — and describes his current mode of work bluntly: he no longer does anything resembling traditional engineering, and "only talks to Claude and has it do everything."
The setup: Claude Code, the desktop app on his Mac, with Opus 5.5 on Medium effort as the default model. From there, the pipeline ran roughly like this:
- Download the full Hacker News comment archive from a Hugging Face dataset.
- Filter it down to comments that genuinely recommend a YouTube video.
- Run Sonnet over those comments in a Claude Code Workflow to strip spam and extract the recommendations.
- Keep the survivors — about 39,834 comments, call it 40,000.
- Wrap the 6,454 unique videos they point at in a YouTube-style UI, organized into 150 topics.
No step in that pipeline involved Karpathy writing code. His fingerprint sits in the prompt, the choice of corpus and the verification pass.
His own contribution was deliberately small. He created a Google Cloud API key so the site could fetch YouTube metadata, booted a VPS, and bought the domain name. Then, in his words, he "gave Claude the IP and told it update, deploy, secure, harden and optimize the code." He describes the collaboration as "a little bit like team work."
If you have never driven a terminal agent, our Claude Code tutorial for beginners covers install, plan mode and permissions. HackerTube's pipeline is more ambitious than a first project, but the mechanics are identical: describe the job, let the model fan out the subtasks, verify what comes back.
Could you build your own HackerTube?
Probably, and that is the point of the data export. Karpathy says an export of the site's data is available "to give to your agent to easily create your own version." You would still need the things he handled by hand — a Google Cloud API key for YouTube metadata, a VPS and a domain — but the corpus, which is the genuinely expensive part, is downloadable.
The copycat wave: one pattern, many sites
Within days, Hacker News produced a family of sites built on the same move: take what HN commenters have recommended over twenty years and pour it into a familiar consumer UI.
| Site | What it aggregates | HN points |
|---|---|---|
| HackerBooks | ~16,500 books mentioned by two or more HN commenters | 1,852 |
| HackerTV | HN's top-recommended videos as 24/7 TV channels | 461 |
| Wikihack | Wikipedia articles ranked by HN mentions | 339 |
More followed: HackTok, HackerZon (Amazon products surfaced from HN comments), HackerGames, HackerWords, HackerProducts. Enough appeared that an Ask HN thread asked whether the flood itself is a concern.
Most of the copies are Claude-built too: roansh's HackerBooks was made mostly with Opus 5.5, and dsalzman's HackerTV with Claude Code. The recipe travels well — swap the corpus, keep the loop.
For scale, HackerTube itself drew 2,877 points, outdrawing every copycat in the family. The original outran the imitations.
What it says about agentic coding
Strip the novelty away and HackerTube demonstrates a shift that is already underway: when describing and verifying become the job, a two-hour solo build can ship a polished product to a very large audience. Karpathy named this territory a while ago — "vibe coding" is his coinage — and HackerTube is the idea carried to its logical end. He did not write the code or pick the architecture; he specified the outcome and checked the result.
The unit of work has changed. It used to be the function; now it is the finished artifact plus the review pass.
It is not the first project of this shape we have covered. Clean Your TV is essentially a job description for an agent with guardrail scripts as the rails, and Ponytail is a single SKILL.md file that changes how coding agents write code. The common pattern: a human sets constraints, the agent does the typing.
Be honest about the boundary, though. HackerTube is a one-off content site — low risk, easily verified, no user data at stake. That is agentic coding's happy path, and it proves the cost of a "weekend project" has collapsed to nearly nothing. It does not prove that agents can maintain a production system with real users, uptime requirements and ongoing change.
Picking an agent to drive? Start with our best AI agent tools roundup.
The limitations, honestly
- The recommender is Hacker News's collective taste — tech-skewed, English-language, drawn from one particular community. Strong for talks and documentaries; not a general-purpose video site.
- 6,454 videos is tiny. A few evenings of viewing, not a platform. You will exhaust the topics you care about far faster than on YouTube.
- Metadata depends on Google. Titles and other metadata come from the YouTube Data API, and the videos themselves are served by YouTube.
- The data is exportable; the code's status is unclear. Karpathy says an export of the site's data is available "to give to your agent to easily create your own version," but neither the writeup nor the site states that the source code is open-sourced.
- Longevity is an open question. The site depends on Google's API and Google's videos, and weekend projects are not always maintained. Nothing has happened — both concerns are context, not predictions.
None of this makes HackerTube less interesting. It makes it exactly what it is: a two-hour demonstration, not a YouTube killer.
FAQ
Is HackerTube free?
Yes. No paywall, no ads on the site itself, no login. You only need your own Google Cloud API key if you take the exported data and build your own version.
Does HackerTube need an account?
No. The ForYou recommendations work without any login — your state lives in the local browser session. There is no account to create and nothing to delete later.
How was HackerTube built?
Andrej Karpathy built it in about two hours on a Friday night with Claude Code, using Opus 5.5 as the default model and running Sonnet over HN comments in a Workflow. His only manual steps were creating a YouTube metadata API key, booting a VPS and buying the domain.
Is HackerTube open source?
Not clearly. The site's data is exportable — Karpathy explicitly invites you to hand the export to your agent and build your own version — but neither his writeup nor the site claims that the source code has been open-sourced.
Will HackerTube be taken down?
No takedown has happened, and neither Karpathy nor Google has commented. Because the videos and their metadata come from YouTube and its API, the site's future sits with Google — treat that as context, not a prediction.


