FazBrowse GitHub Viewer | Trending |
URL:
| Home
Tools: [Download Repo ZIP]   [Original HTTPS Page]

browser-tool-calling-protocol/browser-agent: Browser based agent utilities · GitHub

Repository files navigation

btcp-browser-agent

Give AI agents the power to see and control any browser.

A lightweight foundation for building AI systems that need browser access — automation, testing, web agents, or any browser-based workflow.

Why This Package?

AI agents struggle with browsers because:

  • Raw HTML is too noisy (thousands of nodes)
  • CSS selectors break when layouts change
  • No stable way to reference elements across turns

Browser Agent solves this with smart snapshots:

BUTTON "Submit" [@ref:0]
TEXTBOX "Email" [required] [@ref:1]
LINK "Forgot password?" [@ref:2]

One command gives your agent a clean, semantic view of any page. Stable @ref markers let it interact without fragile selectors.

Features

  • Smart Snapshots - Accessibility tree format optimized for AI comprehension
  • Stable Element Refs - @ref:N markers that survive DOM changes within a session
  • Full Browser Control - Navigation, tabs, screenshots, keyboard/mouse
  • 46 DOM Actions - Click, type, fill, scroll, hover, and more
  • Two Modes - Chrome extension (full control) or standalone (same-origin)

Quick Example

import { createClient } from 'btcp-browser-agent/extension';

const agent = createClient();

// Navigate and understand the page
await agent.navigate('https://example.com');
const snapshot = await agent.snapshot();
// Returns: BUTTON "Login" [@ref:0], TEXTBOX "Email" [@ref:1], ...

// Interact using refs - no CSS selectors needed
await agent.fill('@ref:1', 'user@example.com');
await agent.click('@ref:0');

Use Cases

  • AI Assistants - Let LLMs browse the web and complete tasks for users
  • Browser Agents - Foundation for autonomous web agents that research, navigate, and act
  • Automated Testing - Reliable UI tests with stable element refs that don't break on layout changes
  • Web Automation - Form filling, data extraction, multi-step workflow automation
  • Web Scraping - Extract structured data with semantic understanding of page content

Installation

npm install btcp-browser-agent

Usage Modes

Extension Mode (Full Browser Control)

For Chrome extensions with cross-origin access, tab management, and screenshots.

Background Script:

import { BackgroundAgent, setupMessageListener } from 'btcp-browser-agent/extension';

// Option 1: Just set up message routing
setupMessageListener();

// Option 2: Use BackgroundAgent directly for programmatic control
const agent = new BackgroundAgent();
await agent.navigate('https://example.com');
await agent.screenshot();

Content Script:

import { createContentAgent } from 'btcp-browser-agent';

const agent = createContentAgent();

// Take a snapshot
const { data } = await agent.execute({ action: 'snapshot' });
console.log(data.tree);  // Accessibility tree with refs

// Click an element using ref from snapshot
await agent.execute({ action: 'click', selector: '@ref:5' });

Popup (sending commands via messaging):

import { createClient } from 'btcp-browser-agent';

const client = createClient();

// Navigate and interact
await client.navigate('https://example.com');
const snapshot = await client.snapshot();
await client.click('@ref:5');
const screenshot = await client.screenshot();

Standalone Mode (No Extension)

For use directly in a web page (limited to same-origin, no tab management):

import { createContentAgent } from 'btcp-browser-agent';

const agent = createContentAgent();

// Take a snapshot
const { data } = await agent.execute({ action: 'snapshot' });

// Interact with elements
await agent.execute({ action: 'click', selector: '@ref:5' });
await agent.execute({ action: 'fill', selector: '@ref:3', value: 'Hello' });

API Reference

BackgroundAgent (Extension Background Script)

High-level browser orchestrator that runs in the extension's background script.

import { BackgroundAgent } from 'btcp-browser-agent/extension';

const agent = new BackgroundAgent();

// Tab Management
await agent.newTab({ url: 'https://example.com' });
await agent.switchTab(tabId);
await agent.closeTab(tabId);
const tabs = await agent.listTabs();

// Navigation
await agent.navigate('https://example.com');
await agent.back();
await agent.forward();
await agent.reload();

// Screenshots
const screenshot = await agent.screenshot({ format: 'png' });

// Execute commands (routes to ContentAgent for DOM operations)
await agent.execute({ action: 'click', selector: '#submit' });

Multi-Tab Operations

// Open tabs
const tab1 = await agent.newTab({ url: 'https://google.com' });
const tab2 = await agent.newTab({ url: 'https://github.com', active: false });

// Method 1: tab() handle - interact without switching
const githubTab = agent.tab(tab2.id);
await githubTab.snapshot();
await githubTab.click('@ref:5');

// Method 2: Specify tabId in execute
await agent.execute(
  { action: 'getText', selector: 'h1' },
  { tabId: tab2.id }
);

// Active tab stays tab1 (no switching needed)

ContentAgent (Content Script)

DOM automation agent that runs in content scripts or web pages.

import { createContentAgent } from 'btcp-browser-agent';

const agent = createContentAgent();

// Execute commands
const response = await agent.execute({ action: 'snapshot' });

Available Actions

DOM Reading:

Action Description
snapshot Get accessibility tree with element refs
getText Get element text content
getAttribute Get element attribute value
isVisible Check if element is visible
isEnabled Check if element is enabled
isChecked Check if checkbox/radio is checked
getBoundingBox Get element dimensions

Element Interaction:

Action Description
click Click an element
dblclick Double-click an element
type Type text (keystroke by keystroke)
fill Fill input (instant)
clear Clear input value
check Check checkbox
uncheck Uncheck checkbox
select Select dropdown option
hover Hover over element
focus Focus element
blur Remove focus

Keyboard/Mouse:

Action Description
press Press a key
keyDown Key down event
keyUp Key up event

Other:

Action Description
scroll Scroll page or element
scrollIntoView Scroll element into view
wait Wait for element state
evaluate Execute JavaScript

Element Refs

The snapshot action returns element references for stable selection:

const { data } = await agent.execute({ action: 'snapshot' });
// data.tree: "BUTTON 'Submit' [@ref:5]\nTEXTBOX 'Email' [@ref:3]"

// Use refs in subsequent commands
await agent.execute({ action: 'click', selector: '@ref:5' });

Script Injection

Inject custom JavaScript into the page's main world and communicate with it:

// Inject a helper script
await client.scriptInject(`
  window.addEventListener('message', (e) => {
    if (e.data?.type === 'btcp:script-command') {
      const { commandId, payload } = e.data;
      // Handle command and respond
      window.postMessage({
        type: 'btcp:script-ack',
        commandId,
        result: { /* your data */ }
      }, '*');
    }
  });
`, { scriptId: 'helper' });

// Send commands to injected script
const result = await client.scriptSend(
  { action: 'getData', id: '123' },
  { scriptId: 'helper' }
);

Why script injection?

  • Access page-level APIs (fetch with page cookies, window globals)
  • Interact with page frameworks (React state, etc.)
  • Execute code with full page context

Architecture

The package provides a clean separation between browser-level and DOM-level operations:

┌─────────────────────────────────────────────────────────────────┐
│  Background Script (Extension Service Worker)                    │
│  ┌─────────────────────────────────────────────────────────────┐│
│  │ BackgroundAgent                                              ││
│  │  - Tab management (create, close, switch, list)             ││
│  │  - Navigation (goto, back, forward, reload)                 ││
│  │  - Screenshots (chrome.tabs.captureVisibleTab)              ││
│  │  - Routes DOM commands → ContentAgent                       ││
│  └─────────────────────────────────────────────────────────────┘│
└─────────────────────────────────────────────────────────────────┘
                              │
            chrome.tabs.sendMessage
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  Content Script (Per Tab)                                        │
│  ┌─────────────────────────────────────────────────────────────┐│
│  │ ContentAgent                                                 ││
│  │  - DOM snapshot (accessibility tree)                        ││
│  │  - Element interaction (click, type, fill, hover)           ││
│  │  - DOM queries (getText, getAttribute, isVisible)           ││
│  │  - Keyboard/mouse events                                    ││
│  └─────────────────────────────────────────────────────────────┘│
└─────────────────────────────────────────────────────────────────┘

Package Structure

btcp-browser-agent/
├── @btcp/core          # ContentAgent - DOM operations
│   ├── createContentAgent()
│   ├── DOMActions
│   └── createSnapshot()
│
├── @btcp/extension     # BackgroundAgent - Browser operations
│   ├── BackgroundAgent
│   ├── setupMessageListener()
│   └── createClient()
│
└── btcp-browser-agent   # Main package - re-exports both

Capabilities Comparison

Capability ContentAgent (Standalone) BackgroundAgent (Extension)
DOM Snapshot Yes Yes (via ContentAgent)
Element Clicks Yes Yes (via ContentAgent)
Form Filling Yes Yes (via ContentAgent)
Cross-origin Same-origin only Any page
Tab Management No Yes
Navigation No Yes
Screenshots No Yes

License

Apache-2.0

About

Browser based agent utilities

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages


Back | FazBrowse Home | New Git URL