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

[Owl-Listener]

What this is

The first piece of v2's architecture: turning a checkable claim from narrative into evidence. Designpowers asserts designs are "accessible"; this measures it. It's the first "truth-layer" tool — and it's built as an MCP server, so it works with any MCP client: Claude and Gemini (the hard constraint you set).

Why it matters (the "seam")

Most of Designpowers is prose an LLM performs — and most of that should stay prose, because it's judgment (taste, craft). But some claims have a checkable answer, and contrast is the sharpest example. This replaces "the reviewer says it's accessible" with "the Save button is 2.19:1 — FAIL AA," no guesswork. That's the v2 principle in one tool: prose for judgment, real tools for truth.

What's here (mcp-tools/accessibility/)

• wcag.js — pure, zero-dependency WCAG 2.2 logic (relative luminance, contrast ratio, AA/AAA thresholds incl. the 3.0 large-text/UI rule). The truth, deliberately split from the protocol plumbing so it's unit-testable on its own.
• server.js — thin MCP adapter exposing two tools:
  • check_contrast — one fg/bg pair → ratio + AA/AAA pass/fail
  • check_palette — many pairs (a whole screen or a DESIGN.md palette) → results + summary
  • both return human-readable text and structuredContent
• README — wiring instructions for both Claude (.mcp.json) and Gemini CLI (settings.json) — identical command, because nothing is Claude-specific.

Model-agnostic by construction

This is the proof that v2 can satisfy "must work on Gemini and Claude": the tool speaks MCP, which both clients support. The orchestration-engine choice (Claude SDK vs. LangGraph vs. per-runner) is still open and deferred — but the valuable truth-layer is portable today.

Verification (16/16, and it really speaks MCP)

• 11 pure-logic tests — verified against WCAG reference values (black/white = 21:1) and the exact by-hand numbers from our earlier smoke-tests (Save grey = 2.19, FAIL).
• 5 live MCP round-trip tests — a real MCP client launches server.js over stdio, lists the tools, and calls them, asserting correct evidence. This proves it speaks the protocol, not just that the file parses.
• CI added (check-mcp-tools.yml); repo nets (consistency, routing) still pass; node_modules gitignored, lockfile committed.

Also

• Wired design-md's accessibility overlay to measure via this tool rather than estimate contrast from hex.

Honest notes

• This is additive and changes no existing user behavior — the tool is opt-in (wire it into your MCP config to use it).
• It covers the checkable slice of accessibility (contrast). ARIA/focus/target-size are future tools; "is the hierarchy clear" stays judgment. The win is that the contrast slice stops being faked.
• I verified the MCP round-trip locally; CI will confirm it on a clean runner (first run of this workflow).

---

Generated by Claude Code