Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

8.4 KB
Raw Blame

Vendored design libraries

Three complete MIT source trees live in third_party/: the ThreeUI Community component catalog and the Design+Code agent skills collection (both from Meng To), and image-blaster, an image-to-3D-environment pipeline. This page is the three.ws-side guide to them: what is in each, what it is good for here, and how to actually run one.

The vendoring contract itself (why they are byte-identical, which guards skip them, how to re-sync) is third_party/README.md. Read that before you change anything under third_party/. The short version: never edit a file in place, and never import from third_party/ at runtime.

Why these are in the repo

three.ws is a Three.js platform whose product surfaces are galleries of interactive 3D things, and whose agents are driven by portable SKILL.md folders. These are the two best-maintained open-source collections of exactly those two artifacts, both MIT, both by the same author.

  • ThreeUI answers "what should a live component catalog feel like, and what do 103 finished WebGL components look like when someone else already built them."
  • The skills collection answers "what does a mature public agent-skill library look like," and gives 132 ready workflows for design, web, and game work that our own agents can load today.

third_party/threeui

@designcodeio/threeui 1.2.0. The login-free Community edition of threeui.com: 50 parent components across 111 routes, 141 free variant records, plus the catalog application that renders them.

third_party/threeui/
  src/package-components/   103 published components, one .ts per component
  src/shaders/              48 shader and scene studies (the raw HTML sources)
  src/components/           the catalog shell: browse grid, controls, source tabs
  packages/                 the publishable @designcodeio/threeui package + CLI
  public/                   built demo pages and the catalog's source-code index

Run it

The tree carries its own dependency set and is not an npm workspace of this repo, so install inside it:

cd third_party/threeui
npm install
npm run dev

npm run build:site runs the upstream typecheck, public-boundary audit, Vite build and build audit together, which is the check to run if you ever sync the tree and want to know the copy is intact.

Use a component

The library is published, so a three.ws surface that wants one takes the npm dependency rather than reaching into third_party/:

npm install @designcodeio/threeui
import { AtTheHorizon } from "@designcodeio/threeui";
import "@designcodeio/threeui/style.css";

export function Hero() {
  return <AtTheHorizon />;
}

Our own frontend is vanilla JS modules rather than React, so the more common path here is the second one: open the matching study under third_party/threeui/src/shaders/<name>/ (each is a self-contained HTML document with its own shaders and scene setup), read how the effect is built, and implement it in the three.ws surface that needs it. Keep the MIT attribution with anything you carry across.

What is worth studying first

  • src/components/ is a working answer to the gallery problem our /gallery, /examples and /marketplace pages all solve separately: one shell that gives every catalog entry a live renderer, a controls panel, a variant picker and a copyable source tab.
  • src/shaders/bookshelf/, src/shaders/japanese-tower/ and src/shaders/landscape/ are complete Three.js scenes in single documents, which is the same shape as our embeddable avatar demos.
  • packages/ shows a component library published from inside a catalog repo: relevant every time we graduate a surface out of this monorepo (see package-extraction.md).

third_party/designcode-agent-skills

132 SKILL.md folders in the same Agent Skills format as .agents/skills/, each with its references, scripts, and a runnable demo/.

Group Skills What it covers
web-design 88 scroll storytelling, GSAP ScrollTrigger, marquees, orbit controls, particle fields, layout systems, animation debugging
game-development 20 playable Three.js systems, gameplay QA, game architecture
codex 19 the capture-to-prompt workflows: video to super prompt, HTML to interaction prompts, stitched full-page capture, daily UI inspiration
ui 3 design-first UI prompting
media 2 image sourcing for a build

Load one into a session

A skill folder is portable by design, so pointing an agent at the path is enough:

cat third_party/designcode-agent-skills/agent-skills/codex/video-to-superprompt/SKILL.md

To make one available to Claude Code as a first-class skill on this machine, copy the folder (not a symlink, so the skill and its references travel together) into your skills directory:

cp -r third_party/designcode-agent-skills/agent-skills/codex/video-to-superprompt \
      ~/.claude/skills/video-to-superprompt

They are deliberately NOT copied into .agents/skills/. That directory is the three.ws skill pack we publish: scripts/build-skills-pack.mjs generates skills-pack.json and SKILLS.md from it, the three-ws-core plugin ships it, and workers/okx-chat-bot copies it into its image. Mixing a third party's 132 skills into our published pack would misattribute them and add 90 MB to that worker's build context. Read docs/agent-skills.md for how our own pack is authored.

What is worth studying first

  • agent-skills/codex/video-to-superprompt/ is the flagship: it turns a screen recording of an interface into a prompt detailed enough to one-shot the HTML. Directly useful for building three.ws pages from a reference.
  • agent-skills/game-development/ maps closely onto the /play world and the multiplayer surfaces, both of which are Three.js gameplay systems.
  • Every skill's demo/ is the worked example. That is why the repo's unanchored demo/ ignore rule had to be negated for third_party/: a skill without its demo is a description instead of a proof.

third_party/image-blaster

neilsonnn/image-blaster, a Claude Code workspace that turns one image into a 3D environment in a few minutes: a Gaussian splat (.spz) of the static scene from World Labs Marble, GLB/OBJ meshes of each dynamic object from Hunyuan 3D on FAL, and ambient plus per-object SFX from ElevenLabs on FAL. It is the closest open-source reference to our own splat (splat.md) and Forge generation pipelines.

third_party/image-blaster/
  .claude/skills/    image-blast-{project,plate,world,3d,sfx,image-edit,uncover,wildcard}
  .claude/agents/    one subagent per pipeline stage
  .claude/scripts/   FAL queue, Hunyuan/Meshy 3D, nano-banana and gpt-image-2 edit
  app/               React + Vite viewer for the generated worlds

Run it

cp -r third_party/image-blaster ~/image-blaster && cd ~/image-blaster
cp .env.example .env     # WORLD_LABS_API_KEY and FAL_KEY
claude                   # then: put an image in input/ and ask it to "blast it"

Run it from a copy, not in place: it writes input/ and worlds/ output next to itself, and the vendored tree must stay byte-identical to upstream.

What is worth studying first

  • .claude/skills/image-blast-plate/SKILL.md: the clean-plate step, which removes confirmed foreground objects from the source image before the world is generated.
  • .claude/scripts/asset-pipeline/fal-queue.mjs: the shared FAL queue helper (.env loading, submit, poll, download) that the 3D and SFX stages reuse.
  • app/src/utils/worldLoader.ts: how the viewer discovers generated worlds and resolves each world's local splat URL.

Licence

All three trees are MIT. threeui and designcode-agent-skills are copyright 2026 Meng To; image-blaster is copyright 2026 Neilson Koerner-Safrata. The licence text ships with each tree and must stay with it, and with anything derived from it.

See also