Verify

A verification layer for Claude Code. Reads your spec, runs a browser agent against your local dev server for each acceptance criterion, and returns pass/fail with screenshots — before you push. No CI. No infrastructure.

Install

Prerequisites

• Claude Code with OAuth login (claude login)
• Playwright MCP configured (see below)

Usage

Claude Code Skills

# One-time setup — auto-detects dev server, indexes app
/verify-setup

# Run verification against a spec
/verify

/verify-setup auto-detects your dev server port, indexes routes and selectors from your codebase, and writes .verify/config.json + .verify/app.json. No npm install needed.

/verify asks for your spec, reviews it for ambiguities, then verifies each acceptance criterion using Playwright MCP. Results appear inline with screenshots.

Playwright MCP Setup

claude mcp add playwright -- npx @playwright/mcp@latest --storage-state .verify/auth.json --isolated

Restart Claude Code after adding the MCP server.

Debugging failures

After a run, evidence lives in .verify/runs/<run_id>/:

# Browse raw evidence for a specific AC
ls .verify/runs/*/evidence/<ac_id>/

Each AC's evidence directory contains:

• result.json — verdict, confidence, reasoning, steps taken
• *.png — screenshots captured during execution

Architecture

/verify runs as a Claude Code skill using Playwright MCP for browser interaction: