Hypit is an open source markup language that lets Claude Code or Codex clone a short video and render 100 variants from one command

Hypit turns a video into a text workflow that a coding agent can rewrite, then batch-renders variants where you swap the on-camera face, the script, and the B-roll. The 40-second video from @github.signals scrolls the README of hypit-ai/hypit while a voiceover repeats the repo's pitch. The narrator says "Hippit," but the on-screen project name is Hypit.

What the video shows

Every frame is the same GitHub page, scrolled slowly. The header reads "Clone any viral video with AI agents" with the tagline "1 command, 100 variants, 100M views." Badges show Node.js 22.15+, pnpm 10.33, TypeScript 5.9, and a license labeled "Apache-2.0 with conditions." The star count in the recording is 642. As of 2026-09-16 the repo reports about 7.5k stars, so the recording predates its traction.

Below the fold is a two-panel image: SVML markup on the left and a rendered vertical video on the right. The sample is a football tier-list clip where a goth-styled host says "Ronaldo is D tier." The markup uses tags like <script>, <segment id="ronaldo">, and <HOST>, with inline references such as @hair-gel, @trophy, and @greek-god that point to B-roll assets. The install command shown is npx skills add hypit-ai/hypit -g.

How Hypit works

The README describes Hypit as "a language and system" for AI agents such as Claude Code and Codex. You drop in a reference video and the agent produces an SVML workflow covering footage, captions, B-roll, and effects. The key design choice is that everything is anchored to words in the script instead of timestamps, which is what makes the words swappable without breaking the edit.

The install command adds Hypit as an agent skill; on first use the agent sets up the Hypit executable. You do not have to start from a clone. The README says you can begin from templates or describe the video and let the agent write the workflow from scratch.

Cost and what "runs on your machine" means

The narrator calls it "a tiny production studio that runs on your machine." The renderer runs locally and is free. The README states that generation models are optional and a workflow can compile captions, motion graphics, and code-rendered visuals without calling a model, so a video can cost $0.

The example clips in the repo do use hosted models. The UGC tier-list example lists two Seedance 2 Mini 720p A-rolls, one 2K portrait, and ten B-roll images from GPT Image 2, at a total of $1.15. The podcast example is $1.07 and the street interview is $1.09. Those charges go to the model service, not Hypit. You also need your own coding agent subscription.

License is not plain Apache

The badge says "Apache-2.0 with conditions." The actual LICENSE file is a modified Apache 2.0 that permits internal commercial use, client work, and single-tenant deployments, but prohibits running Hypit as a multi-tenant service, commercially redistributing it, or removing Hypit branding from user-facing output without written authorization. The hypit.ai site sells a commercial API tier and a demo booking for those cases. If you plan to wrap Hypit in a product, read the license first.

Claims I could not verify

The narrator says Hypit "swaps out the face." The repo's GitHub description does use the phrase "swap the face," but the README content I fetched does not document a face-swap or lip-sync integration by name, and the examples generate new A-roll with Seedance rather than altering an existing person's face. Treat "swap the face" as "generate a different presenter" until you test it.

The "100 variants in one go" figure is marketing copy from the repo, not a benchmark. Nothing in the README shows a 100-variant batch or its cost.

Key Takeaways

• Hypit is a TypeScript tool that gives coding agents a markup language (SVML) for building short-form video; the video is a README walkthrough, not a demo.
• Edits are anchored to script words instead of timecodes, which is what makes script and B-roll swaps mechanical.
• Install is one line: npx skills add hypit-ai/hypit -g. Requires Node 22.15+, pnpm, and a Claude Code or Codex account.
• The tool is free; generated footage costs about $1 per 20-second clip in the repo's own examples, billed by the model provider.
• The license is modified Apache 2.0 that blocks multi-tenant hosting and redistribution, so it is not permissive in the usual sense.
• Face swapping is claimed in the repo description but not documented in the README I read.

Resources

• hypit-ai/hypit on GitHub -- the repository shown in the video, with README, examples, and the modified Apache license
• hypit.ai -- official site with templates, pricing page for the commercial API, and Discord link