Agentic Browser Proxy

An autonomous AI-powered browser agent that can perform tasks on websites using natural language goals. Built with Playwright and integrated with LLM backends for intelligent decision-making.

Features

• 30+ browser actions - click, type, scroll, extract data, fill forms, and more
• Human-like behavior - realistic typing, random delays, error recovery
• Skill system - pre-configured automation recipes for specific websites
• Profile management - multiple browser profiles for parallel sessions
• Duplicate tracking - avoid repeating already completed tasks
• Persistent memory - learns from past sessions on each domain

Prerequisites

• Python 3.10+
• Chrome/Chromium browser
• Windows, macOS, or Linux

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Start the LLM Proxy

# Terminal 1 - Start the proxy server
python glm_proxy.py

3. Run the Agent

# Terminal 2 - Run the agent
python run_agent.py "https://google.com" "search for python tutorials" 50

Usage

Generic Mode

For general website automation without a skill:

python run_agent.py "<url>" "<goal>" [max_steps]

Example:

python run_agent.py "https://amazon.com" "find shoes under $100" 50

Skill Mode

For websites with pre-configured skills:

python run_agent.py --skill <skill_name> "<goal>" [max_steps]

Examples:

python run_agent.py --skill quora "answer 5 questions about python" 0
python run_agent.py --skill twitter "post a tweet about AI" 0

Using 0 for max_steps enables unlimited running until the skill's rules are met.

Skill System

The .yaser/ directory contains skill configuration files for specific websites. Each skill defines:

• Profiles - Multiple browser profiles for parallel sessions
• Rules - Answer limits, wait times, profile switching behavior
• Selectors - Custom CSS selectors for site-specific elements
• Instructions - AI instructions for that particular website

Creating a Skill

Create a .yaser/<site>.md file with:

site: quora.com
start_url: https://www.quora.com

profiles:
  - name: profile1
    cookies: []
  - name: profile2
    cookies: []

rules:
  answers_per_session: 5
  wait_between_posts_seconds: 30
  switch_profile_after: 5
  ask_user_to_login: true

tracking:
  completed_file: completed_quora.txt

selectors:
  answer_button: "a.q-text"
  submit_button: "button.q-text"

instructions: |
  Focus on unanswered questions with high views.
  Always be helpful and provide detailed answers.

Available Actions

The agent can perform these actions:

Action	Description
click	Click an element by index
type	Type text into an input
select_option	Select from dropdown
scroll	Scroll the page
goto	Navigate to URL
extract	Extract page content
wait	Wait for specified seconds
goal_completed	Mark task as done
goal_failed	Mark task as failed

Skill Actions (when skill loaded)

Action	Description
check_duplicate	Check if URL already completed
mark_completed_item	Mark URL as completed
switch_profile	Switch to next browser profile

Configuration

Edit agentic_browser_v2/config.py to customize:

• MAX_STEPS - Default max steps per session
• LONG_RUN_MODE - Enable unlimited steps
• VISION_FAILURE_THRESHOLD - Screenshot capture threshold
• YASER_DIR - Skill files directory

Project Structure

.
├── agentic_browser_v2/     # Core agent modules
│   ├── agent.py           # Main agent logic
│   ├── actions.py         # Browser actions
│   ├── prompts.py         # LLM prompts
│   ├── skill_loader.py    # Skill system
│   ├── profile_manager.py # Profile management
│   └── duplicate_tracker.py
├── glm_proxy.py           # LLM proxy server
├── proxy_api.py           # REST API server
├── run_agent.py           # Agent launcher
├── .yaser/                # Skill configurations
└── requirements.txt       # Python dependencies

License

MIT