Hypit
An agent skill that turns a coding agent into a video cloner: npx skills add hypit-ai/hypit -g, then in claude-code or Codex, "Clone this video: /path/to/video.mp4, and
replace…". The README's pitch is the whole workflow, not a script — "swap the face, the words, the
B-roll, ship 100 variants in one command." Read from the README and LICENSE; arrived by Telegram
2026-09-17. 7.5k stars, 893 forks, 1,300 commits at read time.
How it works
- A video is described in SVML, a markup language for compositions in which captions, B-roll and effects are anchored to words, not timecodes — so swapping the script or the language reflows the timing. WhisperX gives word-level alignment; captions are speaker-aware.
- Visuals are code-rendered in headless Chromium, many in parallel (the examples use 64 processes), so a variant costs render time rather than a generation call. Generation is called only for the parts that need it: Seedance 2 Mini for video, GPT Image 2 for stills, Google Video Intelligence and a YOLOv8 anime-face model for detection.
- Three worked examples put a price on a clip: a 20-second "debate" at $1.15, an 18-second podcast at $1.07, a 26-second interview at $1.09 — all API spend, since Hypit itself is free and "your Coding Agent and model services have their own accounts and charges."
- TypeScript monorepo (Node 22.15+, pnpm), packages/services/skills; a hosted "HypiHub" is named.
The licence, read
Called an open-source licence; it is a modified Apache 2.0 that fails the OSI definition. Permitted: internal commercial use, single-tenant self-hosting, forks under the same terms. Forbidden without written permission: running it as a multi-tenant service, selling or bundling it for third parties, and removing the logo from user-facing surfaces. Output belongs to the user. This is the third video skill here and the third distinct licence shape: open-edit is Apache with PolyForm-Shield renderer binaries, video-use is MIT, Hypit is source-available with a no-competing-service clause.
What it adds here
The video-skill row now has three members and a gradient. video-use edits by transcript and composites at decision time; open-edit gives an agent a headless editor; Hypit starts from a finished video and treats it as a template — which is the "clone a viral video" business, and the README says so in its first sentence. Word-anchored composition is the technical idea worth keeping: it is what makes localisation and variant generation a reflow rather than a re-edit. The costs are the vendor's own examples; the 100-variants claim is a capability, not a measurement.
Runner-up spoke: llm-providers (the generative-media cluster — Seedance and GPT Image 2 are called from here). Routed here because the object is a skill for coding agents.
Tier T3 — first-party README and licence; star count unverified elsewhere.