Turn any video format into a style your agent can reuse.
Point your coding agent at 5–20 videos of one format. video-dna writes a style file: hook, shot lengths, captions, music pacing, packaging. shorts-qa checks every render before you post. Everything runs on your machine, with your agent and your keys.
$ npx skills add yourname/video-dna
Install this skill for me: github.com/yourname/video-dna
$ video-dna extract ./refs --out style.json --json
$ shorts-qa check final.mp4 --json
Two tools, one loop
Learn the format, make the video with any renderer, check it, fix it. The agent does the loop; you approve the result.
video-dna
Reads transcripts and a few frames per shot, not whole videos, so your agent spends few tokens. Formats only, never a real person's likeness.
{
"format": "kinetic-lyric-9x16",
"hook": "line on screen by 0.4 s",
"median_shot_s": 0.62,
"cuts_on_beat": 0.91,
"captions": "2 lines, center, word pop"
}
shorts-qa
Deterministic checks with no AI model: pace, dead air, loudness, caption safe zones, first-2-second motion, platform specs.
{
"verdict": "fail",
"loudness_lufs": -19.8,
"dead_air_s": [3.1, 3.9],
"caption_safe_zone": "pass",
"fix": "normalize to -14 LUFS; cut 3.1-3.9 s"
}
Built for agents
Every command speaks JSON and nothing paid runs without a quote.
contract --json tells the agent what each command takes and returns.--dry-run prints minutes and cost before any paid step.Skip the setup. Run it in the cloud.
No Chrome, ffmpeg or GPU on your side. Your agent gets an API key and an MCP endpoint.
Open source
Everything, on your machine.
- video-dna and shorts-qa
- Agent skill, CLI and MCP
- Your own keys for AI clips
- MIT licence
Creator
For one person shipping weekly.
- 120 cloud render minutes
- Cloud shorts-qa with fix list
- Library of format styles
- 1080p, no watermark
Studio
For small teams and agent fleets.
- 600 cloud render minutes
- Agent API key and MCP endpoint
- Batch QA reports
- 3 seats, priority queue
Billed by Paddle. Unused minutes reset monthly. The cloud tier renders and checks only: it does not generate faces or voices and does not post to social media.