What Hypit Actually Is
The pitch is one line: clone any viral video with AI agents. Hypit gives coding agents like Claude Code and Codex a language and a system for making video. Drop in a reference video, and your agent clones it as a complete workflow: the footage, the captions, the B-roll, the effects, all anchored to words instead of seconds.
The distinction the README insists on matters. This is not a script breakdown or a shot-list generator. You get an editable, re-runnable composition, the whole production pipeline as code, which means the clone is a starting point you can remix endlessly: swap the face, the words, the B-roll, ship 100 variants in one command.
And to be clear, cloning is the fastest way in, not the only one. You can start from the included templates, or just describe the video you want and your agent writes the workflow from scratch. Generation models are optional too: a workflow can compile captions, motion graphics and code-rendered visuals into a finished video without calling a generation model or incurring its service charges. That last point is doing quiet work in the background of every example below.
The Core Idea: Video Anchored to Words, Not Seconds
Traditional video editing is timeline editing: everything is pinned to a timestamp. Hypit pins everything to words instead. The composition is written in SVML, the project's markup language for video, where segments, hosts, captions and effects attach to the spoken text.
The payoff shows up the moment you change something. Rewrite a line and the timing re-flows itself, captions re-align, B-roll re-cues. WhisperX handles the word-level alignment, and the examples show speaker-aware karaoke captions driven by face bounding boxes, so the captions know who is talking, not just what was said.
Components are pluggable on top of that. Swap the host without touching the captions. Use the library, fork it, or write your own. The README's framing is deliberate: creating a component is ordinary video-production work, and the same package can be handed off as a versioned tarball or published under your own npm scope. The workflow stays yours and stays editable, which is the opposite of the one-off renders that most AI video tools produce.
The Receipts: Three Videos, About a Dollar Each
The README does something most AI video projects avoid: it publishes the production notes and the cost for three full examples.
UGC: "GOAT DEBATE". A 20-second football tier list putting Ronaldo in D and Messi in S. Two AI-generated A-rolls, a 2K portrait and ten 1K B-rolls, WhisperX word alignment, a sound-synced ranking board, color-box karaoke captions and background music, concurrently rendered in 64 headless Chromium processes. Three clones included: swap the narrator to a banana cat, flip the rankings so Ronaldo becomes GOAT, or replace all players with tech founders. Same viral structure, different viral video. Total cost: $1.15.
Podcast: "DAILY CREATINE". An 18-second podcast clip with split-screen interview layout, speaker-aware captions, a product-handoff moment and background music. Three clones: swap both hosts to Pepe and Doge, replace the creatine with retinol, or ditch the physical product for an app. Same podcast format, three ad verticals. Total cost: $1.07.
Street interview: "NICE RIDE". A 26-second interview with head-tracked speaker-colored captions, a sound-synced emoji reveal board with color flashes and reveal sound effects. Three clones: swap hosts to Wojak and Chad, translate everything to Spanish with the same punchlines, or swap the Lambo for an F1 car. Total cost: $1.09.
The pattern across all three is the product thesis in miniature: one composition, N variants, each a full production, each around a dollar in model-service charges.
What You Can Actually Build With It
The README's use-case list reads like a content agency's service menu, and each one maps to the clone-and-vary loop.
Paid social ads. Clone a winning ad from the Meta Ad Library, swap in your product, ship 50 hook variants the same day. When the creative fatigues in two weeks, re-run with fresh openings; the body never changes.
Viral clones. Any TikTok, Reel or Short becomes a template. Swap the host, the hook, the product, the language, the aspect ratio. If you are reformatting across surfaces, check the destination's preferred dimensions first with our Aspect Ratio Calculator: a vertical Short cropped wrong is just a tall wrong file.
TikTok Shop and affiliate videos. One format that converts, a new SKU every day. Swap the product, the price, the CTA; the structure that worked stays untouched.
AI UGC and talking heads. Narration, word-level captions, B-roll, comment stickers, beat-synced cuts, all wired automatically.
Podcast and interview clips. Split-screen layouts, speaker-aware captions, reaction overlays.
Code-rendered videos. Visuals driven entirely by front-end code, rendered locally without generation API calls, which is where the "generation models optional" claim earns its keep.
Localized versions. The same video in ten languages. Rewrite a line and the timing re-flows itself, which is the word-anchored editing idea paying off directly.
Getting Started: One Install, Then /hypit
Installation is a single command that installs Hypit as an agent skill:
npx skills add hypit-ai/hypit -g
On first use, your agent checks for the Hypit executable and helps prepare it if needed. Your video project can live anywhere. Then you talk to your agent in plain language:
/hypit Clone this video: /path/to/video.mp4, and replace the ranking content with a comparison of Hypit with other AI video products.
Or start without a reference at all:
/hypit Make a ranking video that puts Hypit in S tier.
Your agent checks the environment, requests the credentials the video needs, generates the material and builds the finished composition. Hypit itself is free to use; your coding agent and model services have their own accounts and charges. HypiHub is the recommended hosted model service, but you can also use your own API or local models by telling your agent the service name and its API documentation so it can set up the connection.
Prerequisites are modern but unremarkable: Node.js 22.15+, pnpm 10.33, TypeScript 5.9. There is a quickstart and a development guide at hypit.ai, plus Discord and Telegram communities linked from the README.
The Honest Limitation: License Wording and Real Costs
Two caveats, and both are about reading the fine print rather than the features.
First, the license. The README badge says "Apache-2.0 with conditions", but the LICENSE file itself is titled the Hypit Open Source License, and GitHub reports the license as unasserted. The README is explicit about one part: the videos and other outputs you create belong to you, and third-party models and services may have their own terms. What the "conditions" amount to in practice needs the LICENSE file read, not the badge. If you plan to build a business on this, read it before you ship, the way you would with any source-available project that is not plain MIT or Apache-2.0.
Second, the costs are real even when they are small. The $1.07 to $1.15 figures are model-service charges for the documented examples, billed by the service you choose, not by Hypit. Your coding agent's own usage is separate. The counterweight is genuine: generation models are optional, local models are supported, and code-rendered videos cost nothing in API charges at all. The economics work best when you treat the agent as the director and keep the expensive generation calls to the parts that need them.
The third, softer caveat is youth. The repository was created on 29 July 2026. The documentation, examples and community channels are unusually polished for a two-month-old project, and the launch-partner table suggests real ecosystem momentum, but expect the edges that come with that age.
Repo Health: 19,103 Stars in Two Months
The repository was created on 29 July 2026 and sits at 19,103 stars with 2,150 forks at the time of writing; the latest push landed on 3 October 2026, the day before this page went up. It also holds Trendshift's #1 repository-of-the-day badge. For a project barely two months old, that is not organic discovery, it is a launch that landed, and the star history will tell you whether it compounds or fades.
Good fit if you produce short-form video at volume, ads, affiliate content, localized versions, UGC-style clips, and you want an agent-driven pipeline where one composition becomes a hundred variants without re-editing by hand.
Poor fit if you need long-form editing, you want a no-code tool rather than an agent skill, or you require a plain permissive license grant for commercial use. Read the Hypit Open Source License first, then decide.
