Browser Agent

Browser Agent

Autonomous AI agent that controls a web browser to complete complex multi-step tasks.

The agent uses **Claude** (Anthropic API) with tool use to analyze web pages and take actions — navigating, clicking, typing, scrolling — all autonomously, while you watch in a real browser window.

Features

Browser Automation (Playwright)

• Visible browser window — watch the agent work in real-time
• Persistent sessions — log in manually once, the agent continues with your session across runs
• Numbered element markers overlaid on the page for transparency
• Auto-handling of dialogs, popups, and new tabs

Autonomous AI Agent (Claude + Tool Use)

• Accepts any task in natural language
• Multi-step reasoning with step-by-step execution
• No hardcoded selectors, URLs, or action sequences — the agent discovers everything by exploring

Context Management

• Smart extraction of interactive elements (not raw HTML)
• Screenshots for visual understanding
• Automatic context trimming with action summarization when conversation grows

Advanced Patterns

• Sub-agent architecture — query_dom tool delegates page analysis to a separate Claude call, extracting full page text content and answering questions about it
• Security layer — confirm_action tool requires explicit user approval before destructive actions (payments, deletions, form submissions)
• Error recovery — tracks repeated failures on same actions and prompts the agent to change strategy

Quick Start

1. Install dependencies

npm install
npx playwright install chromium

2. Configure API key

cp .env.example .env
# Edit .env and add your ANTHROPIC_API_KEY

3. Run

npm start

A browser window opens. Type a task in the terminal and watch the agent work.

Usage Examples

Task: Go to google.com and search for "best restaurants in Moscow"

Task: Open hh.ru, find 3 AI engineer vacancies and summarize them

Task: Go to yandex.ru/la