Scroll-World: The Claude Code Skill Behind AI Scroll-Through 3D Landing Pages

Scroll-World is an open source agent skill that builds an AI scroll-through 3D landing page: a site where scrolling drives a camera through a sequence of generated scenes, flying from outside each one into its interior with no cuts between them, in the style of Apple's scroll-through product pages. It runs inside Claude Code, Codex, or any agent that supports the SKILL.md format. The name and the scan description that flagged it suggest something simpler than what it is, so the first thing worth correcting is what "turn a brand into a 3D landing page" actually means here. It does not scrape an existing website and reskin it. It generates an entirely new page from a brief.

What Scroll-World actually builds

Point the skill at a brand and it runs an interview: the industry and pitch, a brand kit (imported from a URL, handed over directly, or proposed by the skill itself), art direction, the ordered list of scenes the camera should visit, whether to render a separate 9:16 mobile version, and a budget. Estimated credit costs for the render tier and art source are shown and approved before anything generates.

From there it produces three kinds of asset per scene: one still image (an isometric diorama, generated through GPT Image 2 via Higgsfield or, optionally, through the Codex CLI's built-in image tool), one "dive-in" camera clip that flies into that scene, and connector clips that join consecutive scenes. The connectors are generated from the actual rendered frames of the neighboring scenes, which is how the seams end up frame-identical instead of cutting. If mobile is opted in, the whole chain renders a second time natively in portrait rather than just cropping the landscape version. The last step wires a config-driven scroll engine that plays the whole sequence as one continuous flight and swaps in the portrait assets automatically on phones. The engine itself is plain, portable JavaScript, so the output drops into a static HTML page, Next.js, Vue, or anything that can serve a video and a scroll listener. Nothing about the output assumes a particular framework.

The accounts and tools it actually requires

This is the part a non-technical reader needs spelled out before getting excited about the demo. Scroll-World is not a hosted app and not a one-click generator. Running it requires:

• Claude Code, Codex, or another SKILL.md-compatible agent to drive the interview and orchestration
• The Monid CLI, with an API key and a funded balance, as the default video-generation backend (Seedance 2.0, billed per clip)
• The Higgsfield CLI, authenticated with its own credits, which renders the scene stills and serves as the fallback video backend if Monid isn't configured
• ffmpeg and ffprobe for frame extraction and encoding
• Python 3 with Pillow, for the mobile portrait canvases and an optional background-knockout step
• Optionally, the Codex CLI, if the user would rather bill scene-still generation to a ChatGPT subscription instead of Higgsfield credits

That is two separate paid AI-generation services to set up and fund, plus command-line tooling, before the skill generates a single frame. There is no version of this where someone without development experience opens a tool, pastes a URL, and gets a 3D site back. It is squarely a builder's tool, run from a terminal, by someone already comfortable installing CLIs and managing API keys across multiple vendors.

What a build actually costs

The README is upfront about pricing rather than hiding it behind "contact sales." A six-scene chain at 1080p through the Monid default backend runs at roughly $27, pay-per-use with no subscription or monthly expiry. The rough cost shape is one image generation per scene plus roughly two video generations per scene minus one, since each connector clip is shared between two scenes. Opting into the mobile version roughly doubles the video generation count, since it renders a second 9:16 chain natively rather than reusing the landscape footage. Higgsfield's own credit pricing isn't exposed through its CLI, so the skill calibrates against whatever balance is already in the account. Either way, the total estimate is shown and has to be approved before any generation starts, which is a reasonable guardrail against a runaway bill.

Who this is actually for

Scroll-World fits a developer, technical marketer, or small studio building a one-off launch page or campaign site who wants a genuinely different visual language than another templated landing page, and who is fine paying somewhere around the price of a decent dinner in generation credits plus their own time wiring up the result. It is a strong fit for anyone who already has Claude Code or Codex in their workflow and has used image or video generation tools before.

It is not a fit for a non-technical small business owner looking for a drag-and-drop site builder. There is no dashboard, no visual editor, and no support flow if a Monid or Higgsfield render fails partway through a chain. It also is not a way to preserve an existing site's actual content, copy, or product photography inside a 3D shell; every asset is generated art, styled as isometric dioramas, standing in for whatever the brand actually looks like. A furniture brand that wants its real product photos flying past the camera will not get that from this skill. What it gets instead is a cohesive, AI-illustrated world built around the brand's pitch, which is a different kind of asset entirely, more suited to a memorable campaign moment than a permanent storefront.

Compared to hiring a WebGL developer to build a genuine scroll-scrubbed 3D scene by hand, which realistically runs into the thousands of dollars and multiple weeks, Scroll-World compresses the same category of output into an afternoon and roughly the cost of a mid-tier subscription, at the cost of giving up hand-modeled control over exactly what appears in each scene.

Traction, as of today

As of today, the repository sits at 8,058 stars and 918 forks on GitHub, carries an MIT license, and lists JavaScript as its primary language. It was created on July 6, 2026, with its most recent code push on July 29 and continued repository activity into this week. Nine issues are open. Star counts measure developer curiosity rather than proof that anyone has shipped a production launch page with it, but the growth is real: an earlier tracking snapshot recorded 6,844 stars, meaning the project has added roughly 1,200 more in a short window.

Installation is a single command inside Claude Code (/plugin marketplace add oso95/scroll-world then /plugin install scroll-world@scroll-world), or through Vercel's npx skills add oso95/scroll-world for Codex and other supported agents, or by manually copying the skill folder from a git clone. The repository is free and open source under MIT, which covers the skill and its scroll engine. The Monid and Higgsfield generation costs sit outside that license entirely and are the real price of admission.