Browser Automation

Browser Automation

Built-in AgentForge Skill — Automate web browsers to navigate, interact, extract data, and capture screenshots.

Overview

The Browser Automation skill gives agents the ability to interact with web pages programmatically using Playwright. This is essential for:

1. Web scraping — Extract text, data, and structured content from any website
2. Form filling — Automate login flows, form submissions, and multi-step workflows
3. Visual verification — Take screenshots for visual QA or documentation
4. Research — Navigate and read web pages to gather information
5. Testing — Verify web application behavior

Supported Actions

How to Use

Setup

import { createBrowserTool, MCPServer } from '@agentforge-ai/core';

// Create browser tool with default config
const { tool, shutdown } = createBrowserTool({ headless: true });

// Register with MCP server
const server = new MCPServer({ name: 'my-tools' });
server.registerTool(tool);

// Or use the convenience function
import { registerBrowserTool } from '@agentforge-ai/core';
const { shutdown: cleanup } = registerBrowserTool(server, { headless: true });

Docker Sandbox Mode

For secure, isolated browser execution (recommended for production):

const { tool, shutdown } = createBrowserTool({
  sandboxMode: true,
  headless: true,
});

This launches the browser inside a Docker container with:

• Isolated network and filesystem
• 2GB shared memory for stability
• Automatic cleanup on shutdown

Direct Import

import { createBrowserTool } from '@agentforge-ai/core/browser';

Agent Instructions

When a user asks you to interact with a web page:

1. Navigate to the target URL first
2. Wait for key elements to load before interacting
3. Use snapshot to understand the page structure (accessibility tree)
4. Use extractText to get readable content from the page
5. Use click and type to interact with forms and buttons
6. Use screenshot to capture visual state when needed
7. Always close sessions when done to free resources

Tips for Reliable Automation

• Prefer #id selectors over class-based selectors
• Use waitForSelector before clicking or typing
• For SPAs, wait after navigation for content to render
• Use extractText with a selector to get specific section content
• Take screenshots before and after critical actions for verification

Configuration Options

Session Management

Each session gets its own isolated browser context with separate cookies, storage, and state. Use sessionId to manage multiple concurrent browsing sessions:

// Session A: logged into site X
await tool.handler({ action: { kind: 'navigate', url: 'https://site-x.com' }, sessionId: 'session-a' });

// Session B: logged into site Y (completely isolated)
await tool.handler({ action: { kind: 'navigate', url: 'https://site-y.com' }, sessionId: 'session-b' });

Prerequisites

• Playwright: npm install playwright && npx playwright install chromium
• Docker (optional): Required only for sandboxMode