QA Agent

Automated exploratory QA testing for web applications — powered by Playwright and, optionally, LLMs (Claude or GPT-4o).

Point QA Agent at a URL and it explores your application like a real user: clicking buttons, filling forms, navigating with the keyboard, and checking for accessibility issues. Then it reports what it finds. No test scripts to write or maintain.

Need targeted tests? Pass plain-English instructions and an LLM generates custom Playwright steps that run alongside the standard suite.

• Features
• Installation
• Quick Start
• Agentic Testing
• Web Interface
• CLI Reference
• Programmatic Usage
• Test Categories
• Output Formats
• CI/CD Integration
• Architecture
• Development
• Contributing
• Exit Codes
• Troubleshooting
• License

Requires Python 3.10+. Check with python --version.

pip install qa-agent            # standard testing (Playwright only)
playwright install chromium     # required — downloads browser binaries

Optional extras:

pip install "qa-agent[ai]"     # agentic testing (convenience marker — no extra packages needed)
pip install "qa-agent[pdf]"    # PDF reports (adds WeasyPrint)
pip install "qa-agent[web]"    # web UI (adds Flask)
pip install "qa-agent[all]"    # everything above

Agentic testing requires an API key for your chosen provider:

export ANTHROPIC_API_KEY=sk-ant-...   # Anthropic (default)
export OPENAI_API_KEY=sk-...          # OpenAI

playwright install chromium must run once after every fresh install. See Troubleshooting if anything goes wrong.

# Test a single URL
qa-agent https://example.com

# Test multiple URLs
qa-agent https://example.com https://example.com/about

# Crawl and test discovered pages
qa-agent --mode explore --max-depth 2 https://example.com

# Generate reports in a custom directory
qa-agent --output json,markdown --output-dir ./reports https://example.com

# Run via module
python -m qa_agent https://example.com

Pass natural-language instructions and an LLM generates custom test steps that run alongside the standard suite. Supports Anthropic (Claude) and OpenAI (GPT-4o and others). No third-party AI packages are required — all API calls use Python's built-in urllib.

# From a bug report (Anthropic, default)
qa-agent --instructions "The login button does nothing when email is blank" \
  https://example.com/login

# Using OpenAI instead
qa-agent --llm openai --instructions "The login button does nothing when email is blank" \
  https://example.com/login

# From a feature spec
qa-agent --instructions "The 'Remember me' checkbox should be unchecked by default \
  and persist the session across browser restarts." \
  https://example.com/login

# From a file
qa-agent --instructions-file feature-spec.txt https://example.com

1. The LLM receives your instructions and the target URL.
2. It returns a structured plan: summary, focus areas, and custom Playwright test steps.
3. The agent runs those steps on every tested page alongside the standard suites.
4. Assertion failures become findings in the report with the severity the LLM assigned.

If the API call fails (or the key is missing), a warning is printed and the run continues with standard tests only.

# Choose provider (default: anthropic)
qa-agent --llm anthropic --instructions "Test checkout" https://shop.example.com
qa-agent --llm openai    --instructions "Test checkout" https://shop.example.com

# Override model (defaults: anthropic → claude-sonnet-4-6, openai → gpt-4o)
qa-agent --llm openai --ai-model gpt-4o-mini --instructions "Test checkout" https://shop.example.com

# Bypass the plan cache
qa-agent --no-cache --instructions "..." https://example.com

Plans are cached to ~/.qa_agent/cache/ (24-hour TTL). Pass --no-cache to force a fresh API call.

python -m qa_agent web              # http://127.0.0.1:5000
qa-agent-web --host 0.0.0.0 --port 8080  # custom bind

• Configuration form with all CLI options
• Real-time streaming output (Server-Sent Events)
• Stop a running test mid-run
• Browse past sessions grouped by domain
• Session detail: findings table, severity breakdown, screenshot gallery, report downloads

No authentication — intended for local or internal use only.

Output is written to output/ by default. CLI sessions appear in the web UI automatically (JSON is always written).

qa-agent --version
qa-agent --help

qa-agent --mode focused https://example.com   # default — test only given URLs
qa-agent --mode explore https://example.com    # crawl and test discovered pages

qa-agent --auth "user:pass@https://example.com/login" https://example.com/dashboard
qa-agent --auth-file auth.json https://example.com
qa-agent --cookies cookies.json https://example.com
qa-agent --header "Authorization: Bearer token123" https://example.com

{
  "username": "testuser",
  "password": "testpass",
  "auth_url": "https://example.com/login",
  "username_selector": "input#email",
  "password_selector": "input#password",
  "submit_selector": "button[type=submit]"
}

qa-agent --output console,markdown,json,pdf https://example.com
qa-agent --output-dir ./reports https://example.com

Default: console,markdown. JSON is always written regardless of --output (for web UI discovery). Output is organized as output/{domain}/{session_id}/qa_reports|screenshots|recordings.

PDF requires the [pdf] extra. Falls back to Markdown if WeasyPrint is not installed.

qa-agent --screenshots       https://example.com  # on errors
qa-agent --screenshots-all   https://example.com  # every interaction
qa-agent --full-page         https://example.com  # full-page captures
qa-agent --record            https://example.com  # session video

qa-agent --no-headless                  # visible browser window
qa-agent --viewport 1920x1080           # default: 1280x720
qa-agent --timeout 60000                # ms, default: 30000

# Skip standard suites
qa-agent --skip-keyboard      https://example.com
qa-agent --skip-mouse         https://example.com
qa-agent --skip-forms         https://example.com
qa-agent --skip-accessibility https://example.com
qa-agent --skip-errors        https://example.com

# Enable opt-in suites
qa-agent --wcag-compliance    https://example.com

from qa_agent import QAAgent, TestConfig, TestMode, OutputFormat
from qa_agent.llm_client import LLMProvider

config = TestConfig(
    urls=["https://example.com"],
    mode=TestMode.EXPLORE,
    output_formats=[OutputFormat.CONSOLE, OutputFormat.JSON],
    max_depth=2,
    max_pages=10,
    instructions="Verify the password reset flow.",  # optional
    llm_provider=LLMProvider.OPENAI,   # optional, default: LLMProvider.ANTHROPIC
    ai_model="gpt-4o-mini",            # optional, default: None (uses provider default)
)

agent = QAAgent(config)
session = agent.run()

print(f"Pages tested:   {len(session.pages_tested)}")
print(f"Total findings: {session.total_findings}")

for finding in session.get_all_findings():
    print(f"  [{finding.severity.value.upper()}] {finding.title}")

TAB order and focusability · Arrow key navigation · Enter key activation · Escape key for modals · Keyboard trap detection · Focus visibility

Click targets · Hover states · Double-click · Right-click/context menus · Target size (WCAG 2.5.5, 44×44 px min) · Overlapping elements

Required field indicators · Validation feedback · Error message accessibility · Label associations · HTML5 input types · Autocomplete attributes

Alt text · Heading structure (h1–h6) · Link text quality · Color contrast · ARIA usage · Landmark regions · Language attributes · Skip navigation

Console errors/warnings · Network errors (4xx, 5xx) · JavaScript exceptions · Broken images · Broken anchors · Mixed content

Non-text contrast (1.4.11) · Use of color (1.4.1) · Content on hover/focus (1.4.13) · Meaningful sequence (1.3.2) · Input purpose (1.3.5) · Focus visible (2.4.7) · Label in name (2.5.3) · Target size (2.5.5) · Language of parts (3.1.2) · Error identification (3.3.1) · ARIA role/property validation

{
  "meta": {
    "session_id": "a1b2c3d4",
    "start_time": "2024-01-15T10:30:00",
    "duration_seconds": 45.2
  },
  "summary": {
    "pages_tested": 5,
    "total_findings": 12,
    "findings_by_severity": { "high": 2, "medium": 5, "low": 5 }
  },
  "findings": [...]
}

# GitHub Actions example
- name: Run QA Tests
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}   # or OPENAI_API_KEY
  run: |
    pip install qa-agent
    playwright install chromium
    qa-agent --output json --output-dir ./qa-results https://staging.example.com

- name: Upload Results
  uses: actions/upload-artifact@v4
  with:
    name: qa-results
    path: ./qa-results/

Exits with code 1 when critical or high severity issues are found, failing the CI step automatically. See Exit Codes.

Omit --instructions / --instructions-file and the API key env vars if you only need standard tests.

qa_agent/
├── cli.py                   # CLI entry point
├── agent.py                 # Core orchestrator
├── config.py                # Configuration dataclasses
├── models.py                # Finding, PageAnalysis, TestSession, TestPlan
├── llm_client.py            # Anthropic & OpenAI clients via stdlib urllib
├── ai_planner.py            # LLM-powered test plan generation
├── plan_cache.py            # Filesystem cache for test plans
├── testers/
│   ├── base.py              # BaseTester abstract class
│   ├── keyboard.py          # Keyboard navigation
│   ├── mouse.py             # Mouse interaction
│   ├── forms.py             # Form handling
│   ├── accessibility.py     # WCAG accessibility
│   ├── wcag_compliance.py   # WCAG 2.1 AA compliance (opt-in)
│   ├── errors.py            # Console & network errors
│   └── custom.py            # AI-generated test steps
├── reporters/
│   ├── console.py           # Colored terminal output
│   ├── markdown.py          # Markdown report
│   ├── json_reporter.py     # JSON report
│   └── pdf.py               # PDF report (requires weasyprint)
└── web/
    ├── server.py             # Flask app with SSE streaming
    ├── templates/            # Jinja2 templates
    └── static/               # CSS and JavaScript

1. Create testers/my_tester.py extending BaseTester, implement run() -> list[Finding]
2. Export from testers/__init__.py
3. Add a test_my_feature: bool flag to TestConfig
4. Call from agent.py in _test_page()

git clone https://github.com/billrichards/qa-agent.git
cd qa-agent
pip install -e ".[dev,web,ai]"
playwright install chromium

# Unit tests (no browser needed)
pytest -v -m "not integration and not network"

# Integration tests (real Playwright)
pytest -v -m integration --no-cov

# Lint & type check
ruff check .
mypy qa_agent

CI runs unit tests across Python 3.10–3.12 on Ubuntu, macOS, and Windows. Integration tests run on Ubuntu with Playwright. See .github/workflows/test.yml.

1. Fork the repository
2. Create a feature branch (git checkout -b my-feature)
3. Make your changes and add tests
4. Run pytest -v -m "not integration and not network"
5. Open a pull request against main

Code style: Ruff + Black, line length 100.

playwright install chromium

Must run once after every fresh install. Easy to forget in CI.

pip install "qa-agent[web]"

Required for qa-agent-web and python -m qa_agent web.

pip install "qa-agent[pdf]"

Falls back to Markdown silently if WeasyPrint is absent.

No extra packages are needed — LLM calls use Python's built-in urllib. You only need a valid API key for your chosen provider:

export ANTHROPIC_API_KEY=sk-ant-...   # for --llm anthropic (default)
export OPENAI_API_KEY=sk-...          # for --llm openai

If the key is missing or the API call fails, qa-agent prints a warning and continues with standard tests.

Requires 3.10+. Check with python --version.

MIT — Copyright (c) 2026 Bill Richards. See LICENSE.