How to Use Hypit: A Real Setup Guide for Claude Code and Codex (2026)

Hypit went from a brand-new repo in late July to more than 10,000 GitHub stars in about seven weeks, and it has been #1 Repository of the Day on Trendshift. The pitch is simple: hand your coding agent a video you like, and it rebuilds that video as a complete, editable project (footage, captions, B-roll and effects) that you can re-run with a new face, product or language.

This guide covers what Hypit actually is, what you need before you start, how your first video runs, what it really costs, and where it fits. Everything here comes from the project's own README, docs and license.

What Hypit actually is

Hypit (github.com/hypit-ai/hypit, site hypit.ai) is an open-source video production system you operate through an AI coding agent such as Claude Code or Codex. It has two parts:

• A skill that gives your agent production knowledge: how to study a reference video, plan the material and review a composition.
• An executable (the @hypit/hypit package) with the commands, components, runtime and a browser-based editor called Studio.

Your project lives in ordinary files. The video is described in .svml, Hypit's markup for the script, media, components and composition, and every run keeps its outputs so you can reuse them. Captions, graphics and effects are rendered in code and anchored to words rather than seconds, so rewriting a line re-flows the timing around it. Generated clips and images come from whichever model services you connect.

You don't have to start from a reference video, either. You can start from one of its templates, or describe the video you want and let the agent write it from scratch.

What you need before you start

• A coding agent that can use skills. Claude Code and Codex are the common choices; the docs also list OpenAgents as an entry partner.
• Node.js 22.15 or newer for the executable. You don't need to download the repo. The agent checks for an existing install and helps prepare anything missing.
• An account for the models, or local ones. The framework is free; hosted generation models are billed by their services (local models and code-only videos avoid that). HypiHub is the recommended hosted service (image, video and voice models plus hosted WhisperX in one account). You can also bring your own keys: the official distribution ships providers for TokenDance, HiAPI, Pollo and Monid, which between them cover models such as Seedance 2.0 and 2.5, MiniMax H3, GPT Image 2 and Nano Banana.
• Word timing for speech. If your reference has talking in it, Hypit uses WhisperX to line words up with the picture, either on your machine or hosted.
• A reference video (a file, or a link from a supported platform) plus any face, product or logo you want in the new version.

Install

In a terminal:

npx skills add hypit-ai/hypit -g

That installs the skill globally for your agent. Then open your video project folder in Claude Code or Codex. On first use the agent looks for the hypit executable and helps install it. To check an existing install later, ask the agent to run hypit version --check. The skill and the executable update through separate channels, so update both when a new release lands.

Your first video, step by step

1. Give it a reference and say what should change. This is Hypit's own example prompt:

/hypit Use this video as a reference: /path/to/video.mp4.
Replace the product with Hypit (hypit.ai), keeping the energetic opening and ranking format.

You can change the presenter, product, language, aspect ratio or call to action. The agent studies why the reference works, including how captions and graphics land on particular words, and writes that reading into the project's reference notes.

2. Pick the services. The agent checks what's already available and explains what's missing. For a spoken reference, that usually starts with choosing local or hosted WhisperX.

3. Agree on a budget. Before any paid generation, the agent lays out the account, the planned work and the estimated cost. You approve a scope, and calls inside that scope go ahead without asking each time. A change of account, scope or budget comes back to you as a new decision.

4. Let it build. The agent prepares the script, performance direction and image references, generates the clips, and builds the captions and graphics while generation runs.

5. Review and edit. It checks layout and timing against the actual material, delivers the video, and can open Studio so you can scrub the timeline, change the properties a component exposes and leave timestamped comments.

The commands under the hood

You rarely type these yourself, but it helps to know what the agent is running:

hypit runtime use hypit.runtime.json
hypit plan build.svrun
hypit build build.svrun --follow
hypit get <build-id> --output final.video --to output/final.mp4
hypit studio --run build.svrun

Only build submits work, including any paid generation. plan shows what would run, get exports the finished file, and studio opens the composition in your browser.

Making variants

This is where Hypit is strongest. Once a video exists as a project, a variant reuses the composition and the existing material and only generates the parts that change. The README's examples each ship three variants of one reference: a swapped narrator, a flipped ranking, a different product category, a Spanish version with the same punchlines. For ad testing (new hooks, a new SKU every day, the same structure underneath) that reuse is the whole point.

What it really costs

• Hypit itself: free. No seat price, no per-render fee, no watermark.
• Models: billed by whichever service you connect. The three README examples, 18 to 26 seconds each, list a total cost of $1.07, $1.09 and $1.15.
• Your agent: your Claude Code or Codex usage is part of the bill. Variants reuse the composition and existing material and only generate the parts that change.

Using Hypit with Claude Code or Codex

Hypit is built around the agent, so this is the normal way to use it, not an add-on. Install the skill once with npx skills add hypit-ai/hypit -g, open any project folder in Claude Code or Codex, and start with /hypit plus your request. If your agent runs on a remote machine, upload the material to that environment and use its preview forwarding: a localhost address on the remote box isn't a preview address on your laptop.

The license, in plain terms

Hypit uses a modified Apache 2.0 license. You can use it for your own organization's work, including commercial work and work for clients, and the videos you make belong to you. What needs a commercial license from Hypit.AI: running it as a hosted, multi-tenant service for other parties, or selling it (or something built from it) as part of a product. The name, logo and copyright notice shown in its CLI, run reports and manifests can't be removed or modified.

Where Hypit fits, and where it doesn't

Hypit is a great fit if you're comfortable in a coding agent, you want precisely timed captions and graphics (rankings, split screens, word-timed reveals), and you plan to push many variants of one format.

It's more work than it looks if you mainly want a finished video from a link. Each new format means a session with the agent that you review and steer, and you choose the model accounts and approve budgets yourself.

If you'd rather skip the setup

Riffkit takes the other side of that trade-off. Paste a winning TikTok link or upload a video, and it studies the formula (the hook, the pacing, the emotional beats) and generates new footage around your product, your character and your language. There's nothing to install and no agent session to run. It generates natively in 9 languages, bills by the second of finished video, and you keep full commercial rights to what you make.

It also runs as an agent skill, so you can stay in Claude Code if you like working that way: see the Claude Code video skill walkthrough. For a side-by-side with Hypit, read Hypit alternative. And if you're comparing open-source agent video tools, How to use OpenMontage covers the other big one.