v2 truth layer: accessibility MCP tool (WCAG contrast) — works on Claude and Gemini

.github/workflows/check-mcp-tools.yml

@@ -0,0 +1,29 @@
+name: Check MCP tools
+
+# Tests the v2 truth-layer MCP tools: the pure WCAG logic and a live MCP
+# round-trip against the server. Runs on PRs and pushes touching the tools.
+
+on:
+  workflow_dispatch: {}
+  pull_request:
+    paths:
+      - "mcp-tools/**"
+      - ".github/workflows/check-mcp-tools.yml"
+  push:
+    branches: ["main"]
+    paths:
+      - "mcp-tools/**"
+
+jobs:
+  accessibility-tool:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: mcp-tools/accessibility
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+      - run: npm install
+      - run: npm test

.gitignore

@@ -0,0 +1 @@
+node_modules/

mcp-tools/accessibility/README.md

@@ -0,0 +1,55 @@
+# Designpowers Accessibility MCP tool
+
+The first **truth-layer** tool for Designpowers v2. It computes **WCAG 2.2 colour-contrast facts** — measured ratios and pass/fail — so an agent gets *evidence* instead of estimating contrast from a screenshot.
+
+It speaks the [Model Context Protocol](https://modelcontextprotocol.io), so it works with **any MCP client — Claude and Gemini alike.** The design knowledge stays portable markdown; this supplies the checkable facts.
+
+## Why this exists
+
+Designpowers' workflow is mostly prose an LLM performs. Most of that is *judgment* (taste, craft) and rightly stays prose. But some claims have a **checkable answer** — contrast is the clearest. This tool turns "the reviewer says it's accessible" into "the Save button is 2.19:1 — FAIL AA," with no model guesswork. It's deliberately split into:
+
+- `wcag.js` — pure, zero-dependency WCAG logic (the truth; unit-tested)
+- `server.js` — a thin MCP adapter over it (the protocol plumbing)
+
+## Tools
+
+- **`check_contrast`** — one foreground/background pair → ratio + AA/AAA pass/fail. `textSize: "large"` uses the 3.0 threshold (≥18pt, or ≥14pt bold, and UI components).
+- **`check_palette`** — many pairs at once (a whole screen or a `DESIGN.md` palette) → per-pair results + a pass/fail summary.
+
+Both return human-readable text **and** `structuredContent` (machine-readable: `ratio`, `passAA`, `passAAA`, thresholds).
+
+## Install
+
+```bash
+cd mcp-tools/accessibility
+npm install
+npm test   # 16 checks: pure WCAG logic + a live MCP round-trip
+```
+
+## Wire it into an agent
+
+### Claude Code / Claude Agent SDK
+Add to your MCP config (e.g. `.mcp.json` or `claude mcp add`):
+```json
+{
+  "mcpServers": {
+    "designpowers-accessibility": {
+      "command": "node",
+      "args": ["mcp-tools/accessibility/server.js"]
+    }
+  }
+}
+```
+
+### Gemini CLI
+Gemini CLI also speaks MCP — add the same server to its `settings.json` `mcpServers` block (the command/args are identical). The tool is model-agnostic by construction; nothing in it is Claude-specific.
+
+## Example
+
+```
+check_contrast { foreground: "#9aa0a6", background: "#e8eaed", label: "Save button" }
+→ Save button: #9aa0a6 on #e8eaed (normal)
+  Contrast ratio: 2.19:1
+  WCAG AA (>=4.5): FAIL
+  Verdict: FAIL (below AA)
+```