Sentinel

Point it at a URL. It explores the app, generates a test plan, runs it, and reports findings. Web + visual regression + accessibility in v0.1; API + mobile in later versions.

Status: alpha, live on PyPI as sentinel-agent==0.1.0a1. Web + visual regression + accessibility ship today. API testing in v0.1.0a2; mobile (React Native) in v0.1.0a3.

Install: pip install sentinel-agent · Repo: GitHub · Issues: file one

What it does

Point Sentinel at a URL:

sentinel run https://your-app.com

In one command, the agent:

1. Opens the URL in headless Chromium
2. Reads the rendered HTML + visible text
3. Asks the LLM to generate a focused test plan (2-5 scenarios, 3-8 steps each)
4. Runs the plan in fresh browser sessions per scenario
5. Captures screenshots and compares against baselines (visual regression)
6. Scans each page state for WCAG 2.1 AA violations (axe-core)
7. Reports findings: failed scenarios, visual diffs, accessibility issues, with cost

Why this exists

The same teams that need Cascade (meeting-to-PR) and Relay (issue-to-PR) need a way to verify that the PRs those agents produce actually work. Hand-writing Playwright tests for every feature is the bottleneck. Sentinel removes the bottleneck: generate tests with the same LLM that writes the code.

Sentinel is fully standalone. It carries its own LLM-client layer and config so it does not depend on any other ThinkNext package at runtime.

Install

# Core install + the LLM provider you want:
pip install 'sentinel-agent[anthropic]'        # Anthropic Claude
pip install 'sentinel-agent[openai]'           # OpenAI
pip install 'sentinel-agent[google]'           # Google Gemini
pip install 'sentinel-agent[claude-code]'      # Local Claude Code subscription, no API key
pip install 'sentinel-agent[all]'              # All providers

# One-time: install the Chromium binary Playwright needs
playwright install chromium

Configure

# Set up an LLM provider. Credentials live at ~/.config/sentinel/config.yaml.
sentinel configure llm anthropic --key sk-ant-xxx --set-default

# Or, if you have Claude Code installed locally (no API key needed):
sentinel configure llm claude_code --set-default

If you want a project-local config (highly recommended; lets you set viewport, baseline directory, accessibility thresholds):

sentinel init

This scaffolds sentinel.yaml with sensible defaults you can edit.

Run

sentinel run https://cascadeagent.dev

# Output (truncated):
#   ✓  3/3 scenarios passed, 0 visual diff(s), 2 a11y violation(s)
#
#   ✓  Homepage loads and primary CTA is visible  (1.42s)
#   ✓  Get-started link navigates to /getting-started/  (1.83s)
#   ✓  Docs sidebar contains all expected sections  (2.10s)
#
#   Accessibility violations:
#     [moderate] color-contrast: Elements must meet minimum color contrast...
#       sample: .text-slate-500
#       (3 node(s) affected)
#     [minor] image-alt: Images must have alt text...
#       sample: img.hero-illustration
#       (1 node(s) affected)
#
#   cost:    $0.04 (5,210 in / 980 out tokens)

What ships in v0.1.0

How it differs from existing tools

Sentinel is for teams who want test coverage without spending the engineering hours to author it. The trade-off is that AI-generated tests have failure modes hand-written tests do not (e.g. an LLM picks a fragile selector). The self-healing v0.1.0a2 feature is the answer to that.

Configuration

sentinel.yaml (after sentinel init):

version: 1

agent:
  provider: anthropic
  model: claude-opus-4-7
  temperature: 0.2

browser:
  headless: true
  viewport_width: 1280
  viewport_height: 720
  timeout_ms: 30000

visual:
  enabled: true
  baseline_dir: sentinel-baselines
  diff_threshold_percent: 0.5

a11y:
  enabled: true
  fail_on:
    - critical
    - serious

Architecture

   sentinel run <url>
          │
          ▼
   ┌──────────────┐
   │ explore page │  Playwright opens URL, grabs HTML + visible text
   └──────┬───────┘
          │
          ▼
   ┌──────────────┐
   │   planner    │  LLM produces TestPlan (2-5 scenarios, 3-8 steps each)
   └──────┬───────┘
          │
          ▼
   ┌──────────────┐
   │    runner    │  Fresh browser session per scenario
   │              │  Each step is one Playwright action
   │              │  screenshot steps → visual regression check
   │              │  a11y_scan steps → axe-core injection
   └──────┬───────┘
          │
          ▼
   ┌──────────────┐
   │ SentinelReport │  Scenarios + visual diffs + a11y violations + cost
   └──────────────┘

Roadmap

License

MIT. See LICENSE.

About

Built and maintained by ThinkNext Software Solutions, alongside our other open-source projects Cascade (meeting-to-PR) and Relay (issue-to-PR).

Follow along: @ThinkNextHQ · LinkedIn · Blog