🌐 AI Web Tester

The Autonomous Eyes & Hands for AI Coding Agents in the Browser

💡 The Missing Piece in AI Coding

AI coding agents can write complex code, refactor components, and fix syntax errors. But there has always been a fundamental blind spot: AI agents cannot see or experience the application running in a real browser.

A login page might have clean code, yet:

• The submit button fails to trigger in specific browser states.

• An unhandled JavaScript exception silently crashes the React hydration root.

• Form validation passes invalid email formats or crashes on edge-case inputs.

• Dynamic modals or infinite scrolls break user navigation.

ai-web-tester bridges this gap. It acts as the high-speed sensory and motor system for AI agents, allowing them to autonomously open URLs (localhost or live), test multi-step user scenarios, catch runtime errors, take visual snapshots of bugs, and compile professional Markdown reports.

Related MCP server: Browser-Debugger

⚡ Key Highlights

1. 🧠 Senior QA Testing Matrices Built-in

Your AI agent doesn't just "click a button" — it executes comprehensive, multi-layer test matrices:

• Empty / Required Field Validation: Checks HTML5 & custom error states when mandatory fields are omitted.

• Format Violations & Boundary Limits: Tests invalid email syntax, password length boundaries, and type constraints.

• Wrong Credentials & State Integrity: Verifies 401 alert banners and error handling without application crashes.

• Security & Injection Guardrails: Evaluates handling of special characters, SQLi patterns (' OR '1'='1), and XSS strings.

• Happy Path (End-to-End Success): Completes full user flows and validates state changes, cookies, and redirects.

2. 🪶 >90% Token Reduction (Semantic State Compression)

Raw HTML DOM dumps consume tens of thousands of LLM tokens and cause selector hallucinations. ai-web-tester prunes the DOM and extracts a clean, numbered Accessibility (ARIA) Tree:

🌐 Current Page: "Customer Portal"
[1] textbox "Email" (required, placeholder="name@domain.com")
[2] textbox "Password" (type=password)
[3] button "Sign In"
[4] link "Forgot Password?"

The AI agent interacts simply by referencing deterministic IDs: click(ref=3) or fill(ref=1, "admin@test.com").

3. 🚀 Lightweight & Headless by Default

Runs entirely in the background using headless Chromium. It consumes minimal RAM and CPU, ensuring your computer stays fast and quiet while your agent audits your application.

4. 📸 Smart Error-Triggered Screenshots

No wasted disk space or clutter. Screenshots are captured automatically only when an anomaly occurs (console error, network 500, unhandled exception, or broken selector) and saved directly into ./artifacts/ with direct clickable links in the report.

5. 🚨 360° Runtime Telemetry & Crash Watchdog

Silently observes browser lifecycle events:

• console.error & console.warn interceptor with precise stack traces.

• Unhandled JavaScript runtime exceptions (pageerror).

• Failed HTTP network responses (4xx / 5xx) and broken asset requests.

• Main-thread freeze & infinite loop detection via watchdog timeout guards.

6. 📄 Automated Markdown (.md) Test Reports

Compiles test results into standardized, human-readable Markdown documents with YAML frontmatter at ./test-reports/report-[runId].md:

• Overall status (PASSED, FAILED, WARNING) and duration metrics.

• Complete step-by-step audit log with ✅ and ❌ status indicators.

• Exact error traces and reproduction steps.

• Embedded links to screenshot artifacts.

• Actionable fix recommendations for the developer or AI agent.

🧙 1-Command Universal Setup Wizard

Easily configure ai-web-tester for any AI agent on your machine with one interactive command:

node dist/cli.js init

Or target a specific agent directly:

# For OpenCode
node dist/cli.js init -a opencode

# For Antigravity
node dist/cli.js init -a antigravity

# For Claude Code / Claude Desktop
node dist/cli.js init -a claude

# For Cursor
node dist/cli.js init -a cursor

# For Windsurf / Roo Code
node dist/cli.js init -a windsurf

# For ALL agents at once
node dist/cli.js init -a antigravity opencode claude cursor windsurf

🤖 Supported AI Agent Ecosystems

🛠️ Model Context Protocol (MCP) Tools

When connected via MCP, your AI coding agent gains access to 6 specialized tools:

┌─────────────────────────────────────────────────────────────┐
│                    AI Web Tester MCP Engine                 │
└──────────────────────────────┬──────────────────────────────┘
                               │
   ┌───────────────────────────┼───────────────────────────┐
   ▼                           ▼                           ▼
browser_open               browser_act              browser_inspect
(Launch & scan DOM)     (click, fill, scroll)     (Refresh state & logs)
   │                           │                           │
   ▼                           ▼                           ▼
browser_screenshot          browser_report           browser_close
(Capture image artifact)   (Export .md summary)     (Clean session teardown)

🖥️ Standalone CLI Usage

You can also run automated browser tests directly from your terminal:

# 1. Quick inspection of interactive elements on any page
node dist/cli.js open http://localhost:3000

# 2. Run automated health check & compile a Markdown report
node dist/cli.js test http://localhost:3000 --title "Homepage Smoke Test"

# 3. Start the MCP Server over stdio manually
node dist/cli.js serve-mcp

📋 Sample Markdown Test Report Output

When your test finishes, ai-web-tester generates a report like this:

---
test_run_id: "rep-1723651234-a8f2"
title: "Authentication Flow - Senior QA Matrix"
target_url: "http://localhost:3000/login"
status: "PASSED"
duration_ms: 3420
total_steps: 5
passed_steps: 5
failed_steps: 0
issues_count: 0
timestamp: "2026-08-14T23:45:00.000Z"
---

# 📋 Laporan Hasil Pengujian: Authentication Flow - Senior QA Matrix

## 📊 Ringkasan Eksekusi
* **Target URL**: `http://localhost:3000/login`
* **Status Keseluruhan**: ✅ PASSED
* **Durasi Pengujian**: 3.42 detik (3420 ms)
* **Total Langkah**: 5 (✅ 5 Passed | ❌ 0 Failed)
* **Total Isu/Error**: 0

---

## 🚶 Urutan Langkah Pengujian (Execution Steps)
✅ **Step 1**: `CLICK` on [ref=3] (button "Sign In") — *Tested empty field validation*

✅ **Step 2**: `FILL` on [ref=1] with value `"invalid-email"` — *Tested invalid email format*

✅ **Step 3**: `FILL` on [ref=1] with value `"' OR '1'='1"` — *Tested security edge cases*

✅ **Step 4**: `FILL` on [ref=1] with value `"valid@company.com"`

✅ **Step 5**: `FILL` on [ref=2] with value `"secretpass123"`

---

## 🚨 Temuan Masalah & Error (Issues & Anomalies)
✅ *No runtime errors, console errors, or failed network responses detected.*

---

## 💡 Rekomendasi untuk AI Coding Agent / Developer
- Semua langkah pengujian berhasil dijalankan tanpa terdeteksi anomali.

🏗️ Architecture & Engineering Principles

Built with clean Hexagonal Architecture (Ports and Adapters):

• Domain Core: Pure domain models (Session, Action, PageState, Issue, TestReport) free from third-party framework dependencies.

• Outbound Infrastructure: High-resilience Playwright driver with smart auto-waiting, auto-scrolling, telemetry observers, and a Handlebars Markdown report generator.

• Inbound Interfaces: First-class Model Context Protocol (MCP) server over standard input/output (stdio) and Commander CLI runner.

• Anti-Zombie Browser Protection: Signal traps on SIGINT, SIGTERM, SIGHUP, and uncaught exceptions guarantee that headless Chromium processes are never orphaned in system memory.

🧪 Development & Contribution

# Install dependencies
npm install

# Download Chromium browser
npx playwright install chromium

# Run full unit & E2E integration test suite
npm test

# Run strict TypeScript typechecking
npm run typecheck

# Build ESM & TypeScript declaration bundles
npm run build

📄 License

Distributed under the MIT License. See LICENSE for more information.