Browse

Analyze any website -- public or auth-gated -- and produce a structured intelligence folder with screenshots, accessibility snapshots, form documentation, and a synthesized Markdown report.

Built on agent-browser.

Quick Start

# 1. Install dependencies
bash scripts/setup.sh

# 2. Analyze a site
bash scripts/site-intel.sh https://example.com

Output lands in output/example-com/. Open site-report.md for the summary.

Prerequisites

Dependency	Install
Node.js	nodejs.org
python3	Ships with macOS. Linux: sudo apt install python3
agent-browser	npm install -g agent-browser (or run bash scripts/setup.sh)

Usage

bash scripts/site-intel.sh <url> [options]

Option	Default	Description
--headed	off	Show browser windows
--max-pages N	10	Max pages to crawl
--output-dir DIR	output/{domain-slug}	Custom output path

Examples

# Quick 3-page scan
bash scripts/site-intel.sh https://example.com --max-pages 3

# Deep crawl with visible browser
bash scripts/site-intel.sh https://example.com --headed --max-pages 50

# Auth-gated site (detected automatically)
bash scripts/site-intel.sh https://app.example.com --headed

# Custom output location
bash scripts/site-intel.sh https://example.com --output-dir ./reports/example

Output

Each run produces a folder:

output/{domain-slug}/
  site-report.md            # Start here -- synthesized report
  sitemap.json              # Discovered pages with titles, URLs, depth
  screenshots/              # Full-page + annotated PNGs per page
  snapshots/                # Accessibility tree JSON per page
  pages/                    # Extracted text content per page
  forms/
    forms-analysis.json     # Structured form field data
    *.png                   # Form screenshots

How It Works

The orchestrator (scripts/site-intel.sh) runs a multi-agent pipeline:

1. Auth Probe        Checks for sign-in pages; opens browser for manual
                     login if detected. Cookies passed to all agents.
         |
2. Navigate          BFS link discovery -> sitemap.json
         |
3. Capture    \
               }     Run in parallel (isolated browser sessions)
4. Forms      /
         |
5. Synthesize        Assembles site-report.md from all collected data

Each agent runs in its own isolated agent-browser --session, so they don't interfere with each other.

Auth-Gated Sites

The pipeline automatically detects sign-in pages before starting analysis. When a login page is found:

1. A visible browser opens so you can sign in manually.
2. Press Enter in the terminal after signing in.
3. Cookies are captured and passed to every agent.
4. The rest of the pipeline runs authenticated.

If no sign-in page is detected, the pipeline proceeds immediately.

Using as an AI Skill

This repo follows the Agent Skills standard. The SKILL.md at the root is a self-contained skill definition that any LLM can load.

To use with Claude Code, copy or symlink SKILL.md into your project's .claude/skills/browse/ directory.

Project Structure

SKILL.md                     # LLM-readable skill definition
README.md                    # This file
scripts/
  site-intel.sh              # Main orchestrator
  setup.sh                   # Dependency checker/installer
  agents/
    navigate-agent.sh        # BFS link discovery
    capture-agent.sh         # Screenshots + snapshots
    forms-agent.sh           # Form element detection
    synthesize-agent.sh      # Report assembly
templates/
  site-intel-report.md       # Report template with {{PLACEHOLDERS}}
.gitignore

Troubleshooting

"agent-browser: command not found" Run bash scripts/setup.sh or npm install -g agent-browser.

Navigate agent finds 0 pages The site may need JavaScript to render links. Try --headed to see what the browser loads.

Screenshots are blank Network-heavy pages may not finish loading in time. Re-running usually helps. You can also increase the timeout via AGENT_BROWSER_DEFAULT_TIMEOUT=30000.

Python errors in synthesize Verify python3 --version works. The synthesize agent uses python3 for JSON processing.

License

Apache-2.0