Open source · MIT · Light-first

The design system for AI-native products

Most design systems were built for forms, tables, and dashboards. Charcoal is built for agent products, on a calm monochromatic backbone with a single black accent.

agent-run · charcoal
PlannerRunning3 tools
+1
  1. Queried the analytics warehouse
  2. Drafted the weekly summary
  3. Waiting on approval to send
Approval required
The agent wants to email the report to 4 recipients.
gpt-tieremail.send
125
components shipped
4
purpose-built layers
200+
design tokens
MIT
open source
Architecture

Four layers, one system

A stable foundation you install, and three fast-evolving AI layers you copy into your repo and own. Every layer reads the same tokens, so it all looks like one product.

Layer 1

Foundations

71 components

The stable base. Installed from npm so every adopter stays in sync.

Layer 2

AI-Native

20 components

The vocabulary of chat and conversation. Copy-in and own it.

Layer 3

Agentic

18 components

Making autonomous work legible: state, cost, permission, and review.

Layer 4

Generative UI

16 components

Interfaces the model assembles at runtime from your data and schema.

Distribution

Install the base. Copy in the rest.

Tokens and foundations are versioned npm packages so upgrades stay in sync. The AI layers are a copy-in registry: run one command, the source lands in your repo, and you fork it freely.

# foundations + tokens, from npm npm install @charcoal/core @charcoal/tokens # copy an agentic component straight into your repo npx shadcn add https://charcoal.nickcoma.io/r/approval-gate.json

Build the agent, not the buttons.

Start from a design system that already speaks the language of AI products.