The substrate is a self-hosted, customer-owned design-system engine
running at design.pointsav.com and at every SMB customer site that
forks it. Its existence is a direct response to four structural gaps
in the 2026 design-system landscape:
Hyperscaler design systems publish only the WHAT. Carbon, Material, Fluent, Polaris, Spectrum, Lightning all publish token values and component shapes. None publish the design-decision research — the WHY — in a form an AI agent or a human at a new organisation can read at codegen time.
SaaS design-system platforms charge enterprise-tier pricing. Specify, Backlight, Knapsack, Tokens Studio Pro target the 5,000+ employee market. The 39% of practitioners working at that scale have options; the SMB beneath them has none.
Every editor wants to be the editor. FIGMA, Sketch, Penpot, Tokens Studio all build to lock the customer into their editor. The W3C Design Tokens Community Group format is the editor- agnostic common denominator the substrate adopts as canonical.
AI codegen is widening the design-system surface area faster than tooling catches up. Only 23% of design systems are structured for AI consumption (December 2025 industry data). FIGMA shipped its design-system MCP June 2025, locked to FIGMA cloud + FIGMA-licensed customers. The first self-hosted MCP-shipping substrate has an 18-24 month structural lead window.
The substrate carries five elements per tenant, in a Git-tracked vault the customer owns:
tokens/ — DTCG-format design tokens. Primitive layer
(Carbon-baselined: color, type scale, spacing, motion, focus
ring), semantic layer (interactive-primary, surface-elevated, ...),
component layer.components/ — HTML+CSS+ARIA recipe files. Framework-
agnostic; the customer's chosen JS/Rust/Elixir/Go framework
consumes the recipe, not the other way around.themes/<brand>/ — per-tenant override layers that re-point
semantic references at primitives.research/ — files like this one. AI-readable design-decision
rationale, accessibility justifications, brand-voice rules,
anti-patterns.exports/ — derived caches (Figma Variables JSON, Tailwind
config, CSS variables, Style Dictionary builds). Recomputable from
the canonical four directories above.The substrate engine (app-privategit-design) reads this vault and
serves it as a public showcase, a DTCG bundle, an AI-readable
research surface, and a Model Context Protocol (MCP) server. AI
agents query the MCP endpoint at codegen time; design tools query
/tokens.json for editor synchronisation; humans read the showcase.
allowed_signers chain. Migration cost falls toward zero — the
customer always has the source.In the AI era of 2026-2030, an SMB's design-system substrate is a medium — its form (machine-readability, editor-agnostic interop, self-hostability, AI-consumable research) shapes how the SMB's brand reaches every customer-facing surface.
The well-structured substrate IS the message the SMB sends to its implementation partners — human or AI. That is the McLuhan position the substrate operationalises.
conventions/zero-container-runtime.md).~/Foundry/CLAUDE.md §6).Read this file first when generating UI for a PointSav-tenant surface. The principles here constrain every subsequent decision:
/tokens.json
endpoint, not your training data's notion of "modern web colors".~/Foundry/DOCTRINE.md §III row 38~/Foundry/conventions/design-system-substrate.md