AI Qadam Video Kit

An AI Qadam Build project.

Make promo, pitch and product videos by talking to a coding agent. You describe the video; the agent writes a brief, lays out the scenes in HTML, records the voiceover, checks the frames and renders an MP4.

The first 18 seconds of videos/kit-promo/, a video about the kit made with the kit. The whole video is 56 seconds long, with voice and music (MP4, 10 MB).

It works with any agent that can read files and run commands (Claude Code, Codex, OpenCode, Cursor, Gemini CLI and others) on macOS, Windows and Linux. The whole stack is open source: rendering by HyperFrames (Apache-2.0), motion by the kit's own runtime (MIT), ffmpeg installed with the packages. What each part does, what is optional and what needs a key: docs/stack.md.

What you get

• A workflow the agent follows: brief with sourced facts → timeline → scenes → checks and stills → voice → render (AGENTS.md, skills/video-kit/).
• One file for timing. video.json holds the scenes, voice lines and sounds, each relative to its scene, so lengthening a scene never desynchronises the voice.
• Product screens drawn in code and acted out: taps, typing, screen changes.
• Voiceover three ways: ElevenLabs; OpenRouter (one key for Microsoft MAI-Voice with native voices per language, Gemini TTS, MiniMax, Fish Audio and more); or free local engines (Kokoro on Apple Silicon, Qwen3-TTS on Apple Silicon, Piper anywhere) with no account at all. Every take is transcribed back and checked against the script. A voice that sounds sharp can be softened in the mix (voiceTone), and its level set against the music (voiceVolume).
• Music: a licensed track (incompetech, by mood and tempo), your own, a track generated from a description (node kit.mjs music: Google Lyria through OpenRouter, several samples to choose by ear, levelled to −18 LUFS), or a synthesised bed; interface sounds generated. The music dips under the voice, and the final mix is normalised to −16 LUFS.
• Media by reference: media search (Wikimedia Commons; incompetech music by mood and tempo; Openverse) and media add (Pexels by clip number, Commons, incompetech, Openverse, any URL with its license): the license and author are recorded, the file is fetched and trimmed, and a preview is made to look at.
• Checks for timing, overlaps, missing translations, uncredited assets, plus HyperFrames' layout, contrast and audio checks; the render reports its loudness.
• Any shape: landscape, portrait, square, or a custom size.

Start

You need Node.js 22 or newer.

git clone https://github.com/aiqadam/video-kit.git
cd video-kit
npm install
node kit.mjs doctor

or, without cloning:

npx github:aiqadam/video-kit init my-videos
cd my-videos && npm install

For a voiceover, copy .env.example to .env and put an ElevenLabs key or an OpenRouter key in it, or install a free local engine: node kit.mjs tts list, then node kit.mjs tts setup kokoro-mlx (native English voices, Apple Silicon), node kit.mjs tts setup qwen3-mlx (Apple Silicon) or node kit.mjs tts setup piper (any OS). Without a key or an engine everything works except the voice: lines are timed by an estimate and stay silent.

Then open the folder in your coding agent and say what you need, in any language:

Make a 40-second vertical video for our coffee shop's Instagram. Here is our menu and logo.

The agent asks for whatever is missing (facts, audience, voice) and builds it under videos/<name>/.

The examples

videos/tutun-lounge/ is a 42-second ad for an invented hookah lounge in Tashkent: the ritual in close-ups (tobacco, bowl, foil, coals, the first draw), cut on the beat of a CC BY track, big type, Russian narration with an Uzbek welcome. The footage comes from Pexels; the kit downloads and trims it on the first build (assets/sources.json).

node kit.mjs render tutun-lounge

The file lands in videos/tutun-lounge/renders/.

videos/kit-promo/ is a 56-second video about the kit, made with the kit. "Qadam" is Uzbek for "step", and the video walks the seven steps of the workflow, showing its own brief, its own video.json and its own frames on screen, then pulls back to reveal the render you are watching. Everything is drawn in HTML except one Pexels clip and a CC BY track; English narration.

node kit.mjs render kit-promo

videos/ottoman-empire/ is a 60-second vertical history explainer for Instagram Reels: six centuries of the Ottoman Empire in ten scenes, from Osman I to the Turkish Republic. No voice: big years that roll forward, Russian captions, a timeline across the top, and paintings, photographs and a map from Wikimedia Commons (public domain, CC0, one CC BY photo), cut on the bars of a CC BY track. Every date in it has a source in its brief.md.

node kit.mjs render ottoman-empire

videos/zero-budget/ is a 58-second English promo about making a promo for nothing: a receipt whose total stays at $0.00 while each scene prints one more line of the bill — the model, the pipeline, the voice, the music, the render. It runs on opencode/mimo-v2.6-flash-free and says so on screen; the voice is Kokoro 82M, generated offline on the machine that rendered it. The terminal, the receipt and the total are drawn in HTML; three stock clips (Pexels, Wikimedia Commons) sit behind them.

node kit.mjs render zero-budget

This is the beginning of the README. Read the rest at the original. Read the full README ↗