Skip to content

Kubai Icons ships an agent skill for animated icons

An open-source repository gives Claude Code and Codex a rulebook, engine and self-check for drawing animated icons, alongside thousands of ready-made React, Vue and SVG icons.

Draftship Editorial3 min read

The kubai-icons repository had 818 stars and 39 forks on GitHub as of 2026-11-14, eleven days after it was created on 2026-11-03. It packages the Kubai Icons agent skill, which teaches Claude Code, Codex or another coding agent to draw animated icons in one consistent style, alongside a large icon library for React, Vue and static SVG. Main language: JavaScript.

What the repository contains

The skill holds the rules, a drawing engine and a self-check. The repository's own table lists 2,800+ icons: nearly 900 objects plus badged and slashed "off" versions, MIT licensed, packaged for React, Vue, plain JavaScript, web components and static SVG. The short description says 2,600+ instead, and the source does not explain the gap. A website at icons.kubai.design, in 16 languages, lets you search by name, shape or meaning, recolor an icon, copy its code or a ready-made prompt for your agent, download SVG or PNG, and save a collection as a ZIP.

Installation is a prompt, not a package. You tell the agent to clone the repository into its skills folder, ~/.claude/skills in Claude Code or ~/.agents/skills in Codex, then restart it. An agent without skill support can read SKILL.md from the cloned folder. A release zip is the alternative.

How the skill draws

You describe the need in plain words; the skill triggers on icon requests, so naming it is optional. The agent writes one line per feature, pairing an everyday object with a hover behaviour ("a folder opens", "a lock springs open"), draws each icon starting from the closest existing example, then runs a headless self-check that compares sizes, outlines, layers and colors until a checklist passes. To start a set for your own product, the skill copies the engine, gallery and self-check into your project.

The README recommends a model that can read images and says the skill was developed and blind-tested with Claude Opus 5.5; other models, including GPT in Codex, can use it. Every icon is a back piece, optional white content and a front piece, with outlines outside the fill and every gradient, outline and shadow derived from one accent color. Each part carries two keyframes, rest and hover.

What changes for people building apps

Four npm packages slot in: kubai-icons-react for React 18+ and Next.js with server rendering, kubai-icons-vue for Vue 3 and Nuxt, kubai-icons for a framework-free API and web component, and kubai-icons-static for standalone SVG. Animation plays on hover, focus, a tap, a parent element or your own state. Color derives from one accent value through CSS, so icons follow currentColor and recolor without script, and the runtime is described as running only while an icon moves. A MyGo port is maintained separately by yldm-tech.

Licenses and limits

Icons and the four packages are MIT. The skill is source-available under the Kubai Icons License; the website, tooling, tests and docs are all rights reserved. The skill license says icons you make belong to you, commercial use included and without credit, but forbids distributing a modified or translated PROMPT.md or SKILL.md, selling the skill, or presenting it under another name. The self-check needs Node.js 18 or later and Chrome; work on the library and site needs Node.js 20. The skill's own SVG export script uses Chinese icon names and palette names.

Open questions

The source does not say which icon count is current, how long a full set takes an agent to draw, or what the packages cost in bundle size beyond one side-effect-free ES module per icon. This suggests the project suits agents that can inspect screenshots and iterate, which likely means slower, costlier runs on text-only models.

What this means for building apps with AI

A design convention now arrives as a rulebook, an engine and a self-check, so the model can grade its own output instead of waiting for a human to eyeball it. For anyone building with an AI app builder, that pattern is worth watching. Draftship turns a plain description into a working app with a built-in database, sign-in and hosting, and paid plans add source code access and GitHub push. Whatever writes the interface, the checks that keep output consistent still have to live somewhere. So does choosing a model that can read its own screenshots, which Draftship's guide to the AI chat and model choice covers.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.