autonomous-browse

name	autonomous-browse
description	Run an autonomous browser agent to complete a task on a website
allowed-tools	Bash
argument-hint	<url> <task>

Autonomous Browser Agent

Run an autonomous browser agent that uses Claude Sonnet to complete tasks on websites.

Usage

The arguments should be: <url> <task description>

Example:

/claude-chrome-stream:autonomous-browse https://github.com "Find the trending repositories page and list the top 5 projects"

Execution

Parse the arguments to extract URL and task:

cd ${pluginDir} && npm run dev -- run --url "$1" --task "$2" --max-steps 30

How It Works

1. Browser Launch: Opens Chrome with the specified URL
2. Visual Streaming: Captures frames via CDP screencast with intelligent delta detection
3. Sonnet Analysis: Sends frames to Claude Sonnet for visual understanding
4. Action Execution: Executes tool calls (click, type, scroll, etc.) from Sonnet
5. Frame Synchronization: Waits for stable frame after each action before continuing
6. Task Completion: Continues until Sonnet indicates task is complete or max steps reached

Features

• Real-time visual feedback: Efficient frame streaming with 2% change threshold
• Hybrid synchronization: Combines lock-step (stable frame waiting) with optimistic (frame IDs)
• Error recovery: Automatic reconnection on session detachment
• Context management: Maintains conversation history for multi-step tasks

Requirements

• Anthropic API key set in ANTHROPIC_API_KEY environment variable or config file
• Node.js 18+
• Chrome/Chromium installed