Hypit (github.com/hypit-ai/hypit) is an open-source system that gives AI coding agents — Claude Code, Codex, OpenClaw and others — a language (SVML) and runtime to create video. You drop in a reference video and the agent clones it as a complete, editable workflow: footage, word-level captions, B-roll and effects, all anchored to words rather than timestamps. It's installed as an agent Skill via 'npx skills add hypit-ai/hypit -g'.

How is Hypit different from a normal AI clip generator?

Most tools output a one-off render. Hypit produces a re-runnable composition — a workflow you can edit and regenerate. Because captions and effects are anchored to words instead of seconds, you can rewrite a line and the timing re-flows automatically, or swap the host without touching the captions. That's what makes '1 command, 100 variants' practical.

Do I have to use paid generation models?

No. Generation models are optional. A Hypit workflow can compile captions, motion graphics and code-rendered visuals into a finished video without ever calling a generation model. When you do use generation (e.g. for A-roll or B-roll), you bring your own model service — HypiHub, your own API, or local models — and pay that service directly. Hypit itself is free with no seat pricing, per-render fees, or watermark.

What does a finished Hypit video actually cost?

The repo documents real examples with itemized costs. A 20-second football 'tier list' clone with two AI A-rolls, ten B-rolls, word-aligned karaoke captions and music — rendered concurrently in 64 headless Chromium processes — came to about $1.15 in model-service charges. A podcast-style ad clone ran ~$1.07 and a street-interview remix ~$1.09. Those are model/service costs, not Hypit fees.

Performance marketers cloning winning ads and shipping hook variants, creators turning any TikTok/Reel/Short into a reusable template, TikTok Shop/affiliate sellers swapping product and CTA daily, and developers who want code-rendered or localized video pipelines. The catch to weigh: it's explicitly built for high-volume viral/ad cloning, which raises obvious originality and likeness-consent questions worth thinking through before you ship 100 variants of someone else's video.

Hypit: Clone Any Viral Video With AI Agents — a Word-Anchored Workflow, Not a One-Off Render

By Prahlad Menon Published 2026-09-16 3 min read

Most “AI video” tools hand you a finished MP4 and call it a day. Hypit takes a different bet: give the agent — Claude Code, Codex, OpenClaw — a language and runtime for video, and let it build a workflow you can edit and re-run. Drop in a viral clip and the agent doesn’t just transcribe it; it clones the whole production as an editable composition: footage, word-level captions, B-roll, and effects, all anchored to words instead of seconds.

Word-anchored, not timestamp-anchored

The core idea is deceptively simple. Everything in a Hypit composition is tied to words rather than timecodes. Rewrite a line and the caption timing re-flows itself. Swap the narrator and the captions stay put. Translate the whole thing to Spanish and the three-stage reveal still lands on the same beats. That’s the mechanism behind the headline pitch — 1 command, 100 variants — because the parts that stay the same never get touched, and only the parts that change get regenerated.

It’s an agent Skill

You install Hypit into your coding agent:

npx skills add hypit-ai/hypit -g

Then you talk to it like any other skill:

/hypit Clone this video: /path/to/video.mp4, and replace the ranking
content with a comparison of Hypit with other AI video products.

Or skip the reference entirely — “Make a ranking video that puts Hypit in S tier” — and the agent writes the workflow from scratch, requests whatever credentials the render needs, generates the assets, and assembles the final composition. The Skill is compatible with Claude Code, Codex, and OpenClaw, and the repo notes integration with OpenAgents Workspace for orchestrating agents across devices.

The receipts

What makes the project credible is that it publishes real examples with itemized production notes and costs:

“GOAT DEBATE” — a 20-second football tier list (Ronaldo in D, Messi in S). Two Seedance 2 Mini A-rolls, a 2K portrait and ten B-rolls from GPT Image 2, WhisperX word alignment, a sound-synced ranking board, color-box karaoke captions, rendered concurrently across 64 headless Chromium processes. Three clones included. Total: $1.15.
“DAILY CREATINE” — an 18-second split-screen podcast ad with speaker-aware captions and a product-handoff moment. Three ad verticals from one format. Total: $1.07.
“NICE RIDE” — a 26-second street interview with head-tracked speaker-colored captions (driven by Google Video Intelligence + YOLOv8 face boxes) and an emoji reveal board. Total: $1.09.

Generation models are optional throughout — a workflow can compile captions, motion graphics, and code-rendered visuals into a finished video without calling any generation API, which is how these land around a dollar.

Genuinely open, genuinely pointed at ads

Hypit is open source with no seat pricing, per-render fees, or forced watermark; you pay only the model service you choose (their hosted HypiHub, your own API, or local models). The use cases it leads with are unambiguous: clone a winning ad from the Meta Ad Library and ship 50 hook variants the same day, turn any Reel into a template, spin a new TikTok Shop video per SKU.

That power cuts both ways. A tool explicitly designed to mass-clone other people’s viral videos — faces, voices, and all — raises real questions about originality, likeness consent, and platform spam policies. If you’re using it for your own footage and brand, it’s a remarkably efficient pipeline. If you’re cloning strangers’ videos at scale, that’s a decision worth making deliberately, not by default.

For anyone building agent-driven content pipelines, though, Hypit is one of the more interesting primitives to appear lately: it treats video as editable, re-runnable code that an agent can author — not a black-box render.

Repo: github.com/hypit-ai/hypit · Demo: hypit.ai