hig-mcp 0.1.0

Structured Apple HIG design-system MCP: tokens, Liquid Glass constraints, SwiftUI mapping. Prose via sosumi.ai.

hig-mcp — Apple HIG design tokens for AI coding agents

An MCP server that gives Claude Code, Cursor, and any MCP client the Apple Human Interface Guidelines as structured data — real system color values, the type ramp, Liquid Glass constraints, and SwiftUI mappings — instead of prose to misread or stale hex to hallucinate.

The problem every Apple dev (and their agent) has right now

At WWDC25 Apple didn't just ship Liquid Glass — it quietly refreshed the system color palette (HIG changelog, June 9, 2025). systemBlue is not #007AFF anymore. It's #0088FF.

Which means:

• Every LLM's training data is wrong. Ask an agent for iOS colors and it confidently hardcodes the pre-2025 palette.
• The HIG is prose, not data. Apple publishes guidelines as web pages; an agent burns thousands of tokens fetching one, then still guesses the numbers.
• Liquid Glass has rules nobody wrote down in one place — blur budgets, compositing layer caps, contrast measured after blur, the mandatory Reduce Transparency fallback. Agents violate all of them by default.

You end up reviewing generated SwiftUI that looks plausible and is subtly off-spec everywhere.

How hig-mcp solves it

Curated, verified token files live inside the server — offline, deterministic, near-zero tokens — and prose is fetched live from sosumi.ai so guidance is never stale. Ask for color and you get the current post-WWDC25 spec table, not a 2023 memory.

Tool	What it returns	Network
hig_get_tokens	Design tokens by category: color, typography, materials, layout, swiftui, sf_symbols	no
hig_check_liquid_glass	Liquid Glass guardrails + a concrete checklist for your context and platform	no
hig_swiftui	HIG component → the right SwiftUI API + which tokens to apply	no
hig_fetch	Current HIG page as clean Markdown (via sosumi.ai)	yes

How it works

Four tools over stdio (Python, FastMCP). The structured half is plain JSON you own and extend (src/hig_mcp/data/); the prose half is delegated to sosumi.ai's DocC-to-Markdown rendering rather than rebuilt. Every value is provenance-tagged:

• apple-system / apple-hig — Apple-published facts (system colors, type ramp, 44pt hit targets).
• wcag-aa — the 4.5:1 contrast rule.
• figma-effect / community-bestpractice — useful numbers Apple never published, flagged so you know to confirm.
• verify: true — beta-era API names that shift; confirm via hig_fetch or Xcode before shipping.

Honest data beats confident data.

Quick start

pipx install hig-mcp
claude mcp add hig -- hig-mcp

Any MCP client, config form:

{ "mcpServers": { "hig": { "command": "hig-mcp" } } }

hig_fetch targets https://sosumi.ai by default; override with HIG_SOSUMI_BASE.

Built-in call counter

Every tool call appends one JSONL line — timestamp, tool, calling client (from the MCP handshake's clientInfo) — to $XDG_STATE_HOME/hig-mcp/calls.jsonl (override with HIG_MCP_CALL_LOG). So you always know which of your agents actually uses it:

jq -r .client ~/.local/state/hig-mcp/calls.jsonl | sort | uniq -c

FAQ

Why not just fetch developer.apple.com? Prose costs tokens and still doesn't contain machine-usable values. Tokens here are instant, offline, and current.

Does it replace sosumi.ai / apple-docs-mcp? No — it deliberately delegates prose to them and owns only the structured layer they don't serve.

What platforms? iOS / iPadOS today, verified macOS type ramp included; values track the current HIG (last verified July 2026, post-WWDC26).

---

Keywords: Apple HIG MCP server · Human Interface Guidelines · design tokens · Liquid Glass · SwiftUI · iOS 26 · macOS Tahoe · Model Context Protocol · Claude Code · Cursor · AI coding agents