Ask an AI for app screens and you get screens. Ask a developer to build them and the first question is: what are the tokens? Which green, which font sizes, which spacing steps?
TL;DR: when you design app screens or a website with SVG Lab (svglab.app) from ChatGPT or Claude, the project now gets a board called 00 Design system: every colour token with its hex value, every text style, the spacing scale, radii, sizes and the parts (buttons, fields, chips, toggles). SVG Lab builds it from the design tokens the screens use and rebuilds it when they change. Shipped in version 1.65.0 on 13 November 2026.
What is on the board
Everything a developer would otherwise eyedrop out of a mockup:
-
Colour: one swatch per value, labelled with the token names that use it (
bg,surface,accent,accent_soft,success,danger_textand so on) and the hex. -
Type: a table, one row per style (
display,title,heading,body,caption,label,button...), with size / line height, weight, font family and tracking. -
Spacing: the scale in px on a 4 px grid (
2xs2 px up to7xl80 px). - Corner radius, control heights, icon sizes, stroke widths and screen layout values (screen margin, card padding, touch target, top bar, tab bar).
- Parts: primary, secondary, outline and ghost buttons, chips, toggles, checkbox, radio, and text fields including the error state.
If the design has illustrations, SVG Lab also lists the colours they are drawn in, since those are not tokens but you still need them.
It is generated from the tokens, not drawn by the model
This is the part that matters for developers. The board is not a picture the AI drew to look like a style guide. SVG Lab builds it from the same design tokens the layout engine uses for the screens, with the first screen the AI lays out. So:
- The values on the board are the values in the screens.
- The AI makes no extra tool calls for it, so it costs no extra credits.
- When the tokens change, SVG Lab rebuilds the board.
We checked point 3. Our test app used this token change:
{
"before": { "accent": "#2F6B4F", "display": "Fraunces", "sans": "DM Sans" },
"after": { "accent": "#3B4FD8", "display": "Bodoni Moda", "sans": "DM Sans" }
}
The board came back with the blue swatches and Bodoni Moda samples in the same step, and SVG Lab worked out new success and danger shades to sit with the new accent.
A side effect: hand edits on the board get overwritten (the board says so in its header). Change the tokens instead, by asking the AI.
It stays out of your build
SVG Lab marks the board as not a screen. It is skipped by Play in Prototype mode, by the screen list and by the code export, so you will not find a DesignSystem page in your React output.
How other tools handle it
Credit where due, because each tool picks a different home for the system:
| Tool | Where the design system lives |
|---|---|
| SVG Lab | A 00 Design system board in the project, built from the design tokens, rebuilt on change |
| Google Stitch | A DESIGN.md file you export and import between projects; Google opened the format as a draft spec on 21 April 2026 |
| Figma Make | Your existing library, brought in with Make kits (npm packages, library styles, guidelines) |
| Banani | A design system generated from a prompt or references, exported to Figma or code |
Best for getting a design system together with app screens you start from a description: SVG Lab, because the board and the screens come from one set of tokens. Best for reusing a Figma library you already maintain: Figma Make. Best for a rules file your coding agent reads: Stitch's DESIGN.md.
Try it
The SVG Lab MCP is a remote server at https://svglab.app/mcp that houses SVG Lab's Design Engine; you sign in with OAuth, no API key. In ChatGPT it is the SVG Lab plugin, in Claude the SVG Lab connector. Setup: svglab.app/mcp-setup. Every account gets 100 free credits (up to 24 app screens).
How the images were made: a two-screen plant care app laid out with SVG Lab's own editor code at version 1.66.3 (the live version on 17 November 2026), in a local browser with no account, using the same calls the Claude connector makes. The screens are deliberately plain; the board is the subject.



Top comments (0)