kosh

A Claude Code plugin for testing WordPress sites. Run automated functional, performance, accessibility, and AEO audits against any live URL using real browser automation via Playwright MCP.

What it does

Kosh tests a site across four dimensions:

• Functional & design: User journeys, layout consistency, link validation, OpenGraph metadata, content quality
• Performance: Load times, console errors, network failures, mixed content
• Accessibility: WCAG 2.2 Level AA compliance: heading hierarchy, alt text, color contrast, keyboard navigation, form labels, ARIA
• AEO (Answer Engine Optimization): How AI tools like ChatGPT, Perplexity, Claude, and Google AI Overviews discover, parse, understand, and cite the site. 49 signals across 8 criteria, status-based evaluation (pass / partial / fail / N/A), site-type-aware and content-driven.

Each test visits 4–6+ pages, simulates real user behavior, and outputs a structured JSON report. A render script converts the JSON into a self-contained HTML report with color-coded severity, collapsible sections, and inline screenshots. The functional, performance, and accessibility test types can also be merged into one comprehensive report.

Quick start

git clone https://github.com/a8cteam51/kosh.git
cd kosh
claude --plugin-dir .

On first run, Claude Code will ask you to trust this project's settings — accept the prompt. This pre-approves the Playwright browser tools so you don't get prompted for each one during a test.

Then run a test:

/kosh:functional-design https://example.com
/kosh:performance https://example.com
/kosh:a11y https://example.com
/kosh:aeo https://example.com

Testing a site that's in coming-soon mode, password-protected, or private? kosh detects the gate and pauses so you can log in — see Testing a site that isn't publicly reachable.

For detailed setup instructions (including troubleshooting), see the Getting Started guide.

Project structure

commands/        Slash commands (/kosh:a11y, /kosh:aeo, /kosh:functional-design, /kosh:performance)
skills/          Full testing procedures for each command
schemas/         JSON schemas for report validation
scripts/         Report generation and merge scripts
hooks/           Session hook that creates reports/data/ on startup
.mcp.json        Playwright MCP server configuration
.claude/         Project settings and Playwright tool permissions

Report structure

Every test type emits a JSON report with this common frame:

{
  "url": "https://example.com",
  "websiteName": "Example",
  "timestamp": "2026-03-20T10:00:00.000Z",
  "visitedPages": ["https://example.com", "..."],
  "issues": {
    "critical": [],
    "high": [],
    "medium": [],
    "low": []
  }
}

Beyond that frame, each test type has its own shape — see schemas/ for the authoritative per-type definitions:

• Functional & design / Performance / Accessibility (qa-report-functional-schema.json, qa-report-performance-schema.json, qa-report-accessibility-schema.json) — mobile and desktop blocks carry the viewport-specific findings (page metadata, network errors, contrast failures, etc.).
• AEO (qa-report-aeo-schema.json) — criteria block with eight criterion sub-blocks (each containing pass / partial / fail / N/A signal evaluations), a top-level summary with signal counts, plus siteType, siteTypeConfidence, and a technicalNotes.applicableSchemas relevance map.