Video with agents · INEMA fork

Clone the viral video, not just the script

You hand over a video and the agent rebuilds the whole flow as an editable project: speech, captions, b-roll and effects.

Start the guide View on GitHub
Hypit banner: clone the viral
In short

Hypit is an open-source tool that teaches agents such as Claude Code and Codex to produce video. You show a reference video and the agent builds a project that works the same way: the script, the images, the voice, the captions and the animations, all tied to the spoken words. It suits people who make short videos and want quick variations (another presenter, another topic, another language). To run it you need Node, pnpm and FFmpeg; the AI-generated video and image parts use paid services, but animation made in code renders locally at no cost. This is a fork of hypit-ai/hypit, tested on the INEMA server.

What it is

A video becomes a project you can rebuild

Instead of editing on a timeline, the agent writes a text file (SVML) that describes the video. Change the speech and the rest follows.

The six parts of Hypit: reference, SVML script, anchor word, karaoke captions, local render and variations

🎯 Clones the system, not the text

The agent watches the reference, transcribes it with the timing of each word and notes why each cut, image and caption appears where it does.

🔗 Events tied to the word

In the script, @{trofeu} … @{/trofeu} marks when an image comes in and goes out. If the sentence changes or moves to another language, the timing adjusts on its own.

♻️ One base, many variations

Swap the presenter, the hook, the product or the language and render again, reusing the material that did not change.

How it works

From reference to MP4

The agent uses the /hypit skill to read the reference and write the project; the hypit CLI checks and renders it.

Reference video→ Transcript with per-word timing→ Reading why it works→ SVML project→ hypit check→ hypit build→ final.video (MP4)

SVML

An HTML-like file with the script, image/voice/video requests, clips, captions and visual components.

Components

Layout and motion packages: ranking scoreboard, karaoke captions, comment sticker, split-screen podcast.

Runtime

Defines where each step runs: HTML rendered in a headless Chrome plus local FFmpeg, or external generation services.

Prerequisites

What you need on the machine

Nothing is paid for the local test in this guide. Tested on 15/10/2026 on a Linux ARM server (NVIDIA GB10).

Node 22.15+ and pnpm 10

We used Node 24 and pnpm 10.33.

node -v; pnpm -v

FFmpeg

Builds the final video and reads the media.

ffmpeg -version | head -1

Headless Chrome

Draws the frames. On x86 Hypit downloads its own; on ARM point it to the Playwright one (step 4).

npx playwright install chromium-headless-shell
User guide · step by step

From clone to first video, no API

The chat.svml example is an 8 s chat animation drawn only in code. It uses no spoken script, no image AI and no source video.

1

Clone and install

The install only downloads JavaScript packages; it does not download a browser.

git clone https://github.com/inematds/hypit
cd hypit
pnpm install --frozen-lockfile
2

Build the types and the example component

The @example/chat-scene component is what draws the messages.

pnpm build:public-types
pnpm --filter @example/chat-scene build
3

Check the project

check validates the SVML and does not call any service.

node bin/hypit.mjs check examples/semantic-composition/chat.svml \
  --project examples/semantic-composition
# ✓ Source is valid · Outputs 11
4

ARM only: point to the Chrome

Hypit tries to use an x86 Chrome that does not exist on ARM (error MANAGED_PROGRAM_DOWN). Create a runtime with chromePath pointing to Playwright's ARM headless build. On x86, skip this step and use hypit.runtime.json.

# examples/semantic-composition/hypit.runtime.arm.json
{
  "format": "hypit.runtime-local@1",
  "dataRoot": ".hypit/runtimes/local",
  "endpoints": {
    "media.local": { "use": "@hypit/media-local", "config": { "defaultConcurrency": 2 } },
    "html.local": { "use": "@hypit/provider-html-local", "config": { "workers": 2,
      "chromePath": "~/.cache/ms-playwright/chromium_headless_shell-1248/chrome-headless-shell-linux-arm64/chrome-headless-shell" } }
  }
}
# replace ~ with the absolute path of your home folder
5

Render

Here it took about 10 s for 240 frames.

node bin/hypit.mjs build examples/semantic-composition/chat.svrun \
  --project examples/semantic-composition \
  --runtime examples/semantic-composition/hypit.runtime.arm.json --follow
# ✓ Build complete · Build bld_...
6

Export the MP4

Use the build id that appeared in the previous step.

node bin/hypit.mjs get bld_... --output final.video \
  --project examples/semantic-composition --to chat.mp4
# 540×960, 8 s, h264 + aac
7

Ask the agent

Install the skill in Claude Code or Codex and describe the video, with or without a reference.

npx skills add hypit-ai/hypit -g
# inside the agent:
/hypit Clone this video: /path/video.mp4 and swap the topic to AI tools
/hypit Make a ranking video that puts INEMA.CLUB in the S tier
Examples

What came out of our test

Frames from the chat.mp4 rendered on the INEMA server, with no API calls.

Frame at 2.5 s: first chat messages
2.5 s: the messages arrive at the times set in the SVML.
Frame at 6 s: four messages on screen
6 s: the whole scene is an HTML/CSS component that stays editable.

The viral examples from the original project (football ranking, podcast, street interview) cost about US$1 each, with video and images generated by paid services. See the examples/ folder.

Limits

What to know before using it

What we weighed while testing.

💳 Generation is paid

The viral clones use Seedance (video), GPT Image (image) and Fish Audio or ElevenLabs (voice), through your own account on each service. Without them, what remains is local: WhisperX transcription, code animation and FFmpeg.

📜 License with conditions

Modified Apache 2.0: your own use, including commercial, is free. Offering it as a service to third parties or reselling it requires a commercial license, and you may not remove the CLI's name and logo. The videos you generate are yours.

🧩 Yet another engine

It is a complete video engine. If you already use HyperFrames or another engine, you can take the ideas without switching tools.

For INEMA

What we will study from it

Ideas noted for INEMA's video engines. None has been applied yet.

Idea 1
Cue tied to the wordMark in the script where each image comes in and goes out, so the timing follows the speech and the EN/ES translations.
Idea 2
Reading the referenceTheir script for understanding why a video holds attention: each element linked to the word it serves.
Idea 3
Ready-made formatsRanking/tier list, split-screen podcast, street interview, talking head, short drama and narrated demo.
Idea 4
Batch variationsOne base video producing several hooks and languages without redoing what did not change.