MCP server · stdio · 14 tools · 2026

Design intelligence over Model Context Protocol.

ux-skill ships a Model Context Protocol server. Any MCP-capable host — Claude Desktop, Cursor, Windsurf, generic agents — can call the recommender, the linter, and all eleven data manifests over stdio. No plugin install, no IDE lock-in, just a long-running daemon and 14 tools.

Why this exists

Reaching design intelligence used to mean installing a plugin into the specific tool the plugin was written for. Claude Code plugins stay in Claude Code. Cursor rules stay in Cursor. The same recommender, the same linter, the same brand specs had to be re-wired for every host. The MCP server collapses that. One process. One protocol. Every host that speaks MCP gets the engine, identically. The recommender is deterministic, the linter is regex-grounded, and both come straight out of the same Python package as the slash commands — so an agent in a fresh repo can ask for the same Stripe-grade landing or fintech-dashboard system that a Claude Code session can.

The asymmetric position

None of the top eight ship one. We do.

The competition is locked inside Claude Code's plugin runtime. The MCP server makes ux-skill reachable from any agent — even agents that have never heard of a Claude Code plugin.

Plugin	MCP server	Effective reach
ux-skill	yes `ux-mcp` over stdio	Claude Desktop, Cursor, Windsurf, any MCP host, generic agents
ui-ux-pro-max-skill	no	Claude Code only
open-design	no	Claude Code only
taste-skill	no	Claude Code only
huashu-design	no	Claude Code only
stitch-skills	no	Claude Code only
nothing-design-skill	no	Claude Code only
hallmark	no	Claude Code only
material-3-skill	no	Claude Code only

One server. Every host. That's the move.

The catalogue

Fourteen tools, one engine.

Each tool is a pure dict -> dict Python handler. Same code path the slash commands use; same data manifests; same deterministic recommender. The MCP layer is a thin stdio wrapper.

Tool	What it does
`ux_recommend`	Run the 5-parallel-search recommender over a brief. Returns style, palette, type pair, motion presets, components, brand exemplars, anti-pattern guardrails, and a rationale line per lane.
`ux_lint`	Run the 120-rule anti-AI-slop regex linter over file paths. Returns findings with rule id, severity, file, line, excerpt, and fix.
`ux_styles`	Return all 84+ entries from `data/styles.json` — design philosophies with tokens.
`ux_palettes`	Return 176+ entries from `data/palettes.json`. Optional `mode` filter (`light` / `dark`).
`ux_type_pairs`	Return 70+ entries from `data/type-pairs.json` — display + body + mono pairings.
`ux_components`	Return 148+ entries from `data/components.json`. Optional `category` filter.
`ux_industries`	Return 184+ entries from `data/industries.json` with style/palette/type biases per domain.
`ux_motion_presets`	Return 57+ entries from `data/motion-presets.json` with easing, duration, stagger tokens.
`ux_anti_patterns`	Return 35+ entries from `data/anti-patterns.json` (the regex rules that drive the linter). Optional `severity` filter.
`ux_brands`	Return 131 brand specs from `data/brands/*.json` (Apple, Stripe, Linear, Tesla, Notion, etc.). Optional `category` filter.
`ux_landing_patterns`	Return 24+ entries from `data/landing-patterns.json` — proven landing-page section patterns.
`ux_persist_save`	Write a recommendation as `.ux/design-system/MASTER.md` in the given project root. Idempotent — same input writes byte-identical bytes.
`ux_persist_load`	Load `.ux/design-system/MASTER.md` back into a structured dict, or report not-found.
`ux_stats`	Return the engine version and per-manifest entry counts. Useful as a health probe.

Example calls

What it actually does.

ux_recommend — get a complete system from a brief

// Input
{
  "brief": {
    "project_type": "landing",
    "industry": "fintech-neobank",
    "tone": ["warm", "editorial", "precise"],
    "must_have": ["dark-mode", "a11y-AA"],
    "forbidden": ["brutalism", "purple-gradients"],
    "stack": "nextjs-15-app-router",
    "region": "mena"
  }
}

// Output (abridged)
{
  "style": { "id": "monochrome-precise", "name": "Monochrome Precise", ... },
  "palette": { "id": "dark-editorial-cinema", "mode": "dark", ... },
  "type_pair": { "id": "bricolage-inter-jetbrains", ... },
  "motion": [ ... /* top 5 presets compatible with style */ ],
  "components": [ ... /* top 12 compatible components */ ],
  "brand_exemplars": [ ... /* top 5 brands to study */ ],
  "guardrails": [ ... /* all 120 anti-pattern rules — always-on */ ],
  "rationale": [
    "Industry: Fintech Neobank",
    "Style: Monochrome Precise",
    "Palette: Dark Editorial Cinema",
    ...
  ]
}

ux_lint — anti-slop regex sweep across paths

// Input
{ "paths": ["./app", "./components"], "threshold": "high" }

// Output (abridged)
{
  "files_scanned": 142,
  "rules_loaded": 35,
  "exit_code": 1,
  "findings": [
    {
      "rule_id": "generic-placeholder-copy",
      "severity": "high",
      "file": "./app/page.tsx",
      "line": 42,
      "excerpt": "<p>Your headline goes here.</p>",
      "fix": "Write real copy that names the user value."
    },
    ...
  ],
  "summary": { "critical": 0, "high": 4, "medium": 7, "low": 12, "total": 23 }
}

ux_stats — engine health probe

// Input: {} ; Output:
{
  "version": "2.0.0-alpha.1",
  "counts": {
    "styles": 84, "palettes": 176, "type-pairs": 70,
    "components": 148, "industries": 184, "chart-types": 35,
    "tech-stacks": 25, "ux-guidelines": 112,
    "motion-presets": 57, "anti-patterns": 35, "brands": 72
  }
}

Client configuration

Connect any MCP host in three lines.

The server is one binary: ux-mcp. Once installed, every MCP host points to it the same way.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "ux-skill": {
      "command": "ux-mcp",
      "args": [],
      "env": {}
    }
  }
}

Cursor (project + global)

.cursor/mcp.json or ~/.cursor/mcp.json

{
  "mcpServers": {
    "ux-skill": {
      "command": "ux-mcp"
    }
  }
}

Then in .cursorrules: "When designing UI, call ux-skill MCP tools first: ux_recommend for the system, ux_lint to catch slop."

Windsurf

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "ux-skill": {
      "command": "ux-mcp"
    }
  }
}

Generic stdio client

any JSON-RPC over stdio host

# Spawn ux-mcp as a subprocess.
# Write JSON-RPC frames to stdin.
# Read responses from stdout.
# Logs are on stderr only.

$ ux-mcp
# or, equivalently:
$ python3 -m engine.mcp.server

How it's built

Transport separated from handlers.

The 14 tools are pure dict -> dict Python functions in engine.mcp.server. The MCP stdio transport is an optional dependency that lazy-imports inside run_server(). Three consequences fall out:

The engine.mcp module imports cleanly whether or not the mcp Python package is installed.
Tests exercise the handlers directly — no async event loop, no subprocess plumbing in the test suite.
ux-mcp fails fast with a clear pip install hint if the transport library is missing.

One source of truth, one entry point, every host.

Install. Connect. Run.

Three lines to design intelligence in any agent.

MIT licensed. No telemetry. No account. The MCP transport is opt-in so the base install stays slim.

$ pip install 'uxskill[mcp]'

$ ux-mcp # starts the stdio server

— then point your MCP host at ux-mcp —