arXivisual: the tool that turns an academic paper into a 3Blue1Brown style animated explainer
You give it the address of any paper on arXiv, and the tool splits the paper into sections, picks the concepts in each section worth animating, and turns them into narrated Manim animations. It is free, its code is public on GitHub, and was built at Carnegie Mellon's 24 hour hackathon, TartanHacks '26. Two things in the usual pitch are not quite right: you do not change two letters in the address, you add a word; and the output is not one standalone video but a readable page with the animations embedded between the text. Below: the correct way to use it, the agent pipeline behind it, and verified steps if you want to run it on your own machine.
The tool itself
Go to the site and paste an arXiv link or a paper ID. It is free; the site shows no pricing or sign up.
What is extra below: the correct form of the address trick, what the output actually looks like, which agents process the paper and in what order, and which key you cannot skip if you want to run it yourself.
The address trick: not two letters, one word
The usual line is "change two letters in the URL". The accurate version: you append isual to the word arxiv. The rest of the address, meaning the paper ID, stays exactly the same. Using the Transformer paper as an example:
arxiv.org/abs/1706.03762
arxivisual.org/abs/1706.03762The output is not a single video file. The paper becomes a page you read section by section, and as you scroll, narrated animations play at the relevant spot. So the text stays in front of you while you watch the explanation, and formulas render properly. In the developers' own words results arrive within minutes; longer papers take longer.
What happens behind the scenes
The summary "one agent reads, one designs the lesson, one animates it" is right but incomplete. The pipeline in the repo has six stages, and its most important part is that the generated animation code passes four separate checks before it ever ships:
From paper to animation
- 1
Ingest
Pulls the paper from arXiv and splits it into sections
- 2
Section analysis
Works out what each section is about
- 3
Visualization planning
Picks the concepts worth animating; this is the step that actually designs the lesson
- 4
Manim code
Writes runnable animation code together with a storyboard
- 5
Four checks
Syntax, on screen layout, narration quality and a test render; if it fails it retries up to 5 times
- 6
Render
Produces the animation with its voiceover and embeds it in the page
That checking layer is what sets the tool apart. Getting an AI to write Manim code is easy; the hard part is getting an animation where text does not spill off screen, narration matches the visuals, and the thing actually runs.
The claims, one by one
What is said, what is true
Change two letters in the URL
Append isual to arxiv: arxivisual.org/abs/<id>
It hands you a narrated video
It gives you a scrolling page with the animations embedded between the text
It first finds the single hardest concept
It picks the concepts worth animating in each section, not one single concept
Built in one day, beat hundreds of teams
Built at the 24 hour TartanHacks '26 and listed as a winner; over 1,000 people attended, but the team count was never published
If you want to run it yourself
For most people the site is enough. But the code is open; if you want to process papers with your own key, the repo has two parts: a Python backend and a Next.js frontend. The backend needs uv, the frontend needs Node.
git clone https://github.com/rajshah6/arXivisual
cd arXivisual/backend
cp .env.example .env
uv sync
uv run uvicorn main:app --reload --host 0.0.0.0 --port 8000cd arXivisual/frontend
npm install
npm run devA Claude key will not work here
The hackathon version used Claude for its agents, but the current version in the repo runs on Azure OpenAI (GPT-5). Nothing gets generated until AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_KEY and AZURE_OPENAI_DEPLOYMENT are filled in the .env file. Narration also goes through the same Azure account by default; to switch to free Google TTS, set VOICEOVER_TTS_SERVICE=gtts.
There is no license file
The repo declares no license. Reading the code and trying it for yourself is fine, but code with no license is all rights reserved by default. If you plan to build a product on top of it, ask the developers first.
DOA: Yapay Zeka ve Otomasyon
Installing these tools on your own is one thing; actually building with them is another. The community has people using these daily and people building systems from scratch.
If you want a system that actually runs in your business, let's talk for 10 minutes; I'll look at what you're trying to build and tell you which path fits. Free, and not a sales pitch.
Book a 10-minute callIf you'd rather learn this alongside people doing the same work instead of on your own, the community is always open:
Join the communityThis is an affiliate link.