4.8
AGBrowse

AGBrowse

Desktop Appai-automation#opensource5 ReditVerified

Ultimate browser automation toolkit for AI agents. Zero MCP token tax, JS evaluate, vision-click with DPR correction.

SHARE
Screenshots (1)
AGBrowse Screenshot #1
Screenshot 1 of 1

Overview

AGBrowse (agbrowse) is a serverless, ultra-fast Chrome DevTools Protocol (CDP) browser automation, adaptive search, and web-AI CLI designed specifically for AI coding agents. Operating without any persistent MCP server overhead or token tax, AGBrowse executes short-lived Node.js commands that inspect live Chrome DOM snapshots via stable numeric references, perform vision-guided coordinate clicks with DPR correction, escalate URL fetches intelligently, and orchestrate live ChatGPT, Gemini, and Grok web sessions.

How It Works

AGBrowse operates through structured, deterministic execution stages:

Serverless CDP Process Model: Reconnects to Chrome DevTools Protocol (:9222) via short-lived Node.js commands, eliminating the heavy memory footprint of long-running MCP servers.
Ref-Based DOM Observation: Captures clean DOM snapshots and maps interactable elements to stable numerical references for deterministic clicking, typing, and form submissions.
Adaptive URL Fetch Ladder (v2): Escalates URL parsing through DNS-pinned direct fetch, TLS impersonation, yt-dlp media extraction, and headless Chrome rendering with BM25 reranking.
Zero-Token Web-AI Sessions: Drives live ChatGPT, Gemini, and Grok web sessions, generates code-mode ZIP artifacts, and retrieves outputs with zero Model Context Protocol token tax.

Key Technical Features & Capabilities

Zero MCP Token Tax: Eliminates bulky JSON tool schemas from LLM context windows by replacing persistent MCP servers with clean, inspectable terminal commands (agbrowse snapshot, agbrowse navigate).
Stable Ref & DPR Vision Clicking: Maps interactive DOM nodes cleanly across page turns and pairs accessibility snapshots with device-pixel-ratio corrected vision clicking (agbrowse-vision-click).
6-Phase Adaptive URL Fetch & Search: Runs agbrowse fetch and agbrowse search with SSRF DNS-pinning (curl --resolve), TLS impersonation, yt-dlp parsing, RSS/Atom feeds, BM25 reranking, and evidence scoring.
Durable Web-AI Sessions & Code Artifacts: Submits prompts to live ChatGPT Pro / Gemini / Grok browser tabs asynchronously (web-ai send / poll), extracting generated /mnt/data/*.zip artifacts automatically.

Architecture & Technology Stack

Primary Languages: JavaScript (ES Modules, Node.js 18+, JSDoc CheckJS Type-Checked)
Browser & Search Runtime: Chrome DevTools Protocol (CDP Port 9222), Narrow Opt-In MCP Bridge, Vitest Test Suite
Extraction & Hardening Modules: Adaptive Fetch v2 (TLS Impersonation, Feed Parser, BM25 Reranker, Structured HTML Extractor)
Supported Platforms: macOS, Windows 10/11, Linux (with local Chromium / Chrome)

System Requirements

SpecificationMinimum RequirementRecommended Specification
Operating SystemmacOS, Windows 10/11, Linux (Node.js >= 18 + Chrome)Latest stable release build
Processor Architecture4-Core x86_64 or Apple Silicon M18-Core Apple Silicon / Core i7+
Memory (RAM)4 GB Unified Memory16 GB+ High-Speed RAM
Disk Space1 GB free SSD storage5 GB+ free SSD storage

Installation & Setup Guide

5 Redit

Global npm CLI Installation (No Lifecycle Scripts)

bash
npm install -g agbrowse
agbrowse --help
agbrowse start

Source Installation & Verification

bash
git clone https://github.com/lidge-jun/agbrowse.git
cd agbrowse
npm install
npm link
npm run gate:all

Installation & Setup Guide

Mở khóa ứng dụng để xem chi tiết hướng dẫn cài đặt, cấu hình môi trường và vận hành.

How to Use & Workflow

5 Redit
1
Start & Observe Browser: Run agbrowse start, navigate to a target site via agbrowse navigate "https://...", and capture interactive nodes using agbrowse snapshot --interactive --max-nodes 120.
2
Fetch & Verify Evidence: Read any candidate URL reliably with agbrowse fetch "<url>" --json --trace, or score search candidates via agbrowse search "<query>" --json.
3
Run Asynchronous Web-AI Tasks: Dispatch long ChatGPT/Grok/Gemini sessions with agbrowse web-ai send --vendor chatgpt --model pro and retrieve results cleanly via agbrowse web-ai poll.
4
Install Agent Skills: Register bundled browser, web-ai, search, and vision-click skills into your Codex or CLI-JAW workspace using agbrowse skills get core --full.

How to Use & Workflow

Mở khóa ứng dụng để xem chi tiết hướng dẫn cài đặt, cấu hình môi trường và vận hành.

Troubleshooting & Common Issues

Issue: Machine-readable parsing fails on error output inside agent scripts.
*Cause*: Default CLI text errors vary by subcommand.
*Resolution*: Export AGBROWSE_JSON_ERRORS=1 to receive deterministic JSON error envelopes across all commands.
Issue: Adaptive fetch rejects target localhost or private IP URLs.
*Cause*: Built-in SSRF mitigation pins DNS lookups and rejects unsafe per-hop redirects by default.
*Resolution*: Use agbrowse fetch <url> --browser required for authorized local web app debugging via Chrome CDP.

How to Use & Getting Started

Protected
## Installation & Setup Guide ### Global npm CLI Installation (No Lifecycle Scripts) ```bash npm install -g agbrowse agbrowse --help agbrowse start ``` ### Source Installation & Verification ```bash git clone https://github.com/lidge-jun/agbrowse.git cd agbrowse npm install npm link npm run gate:all ```

How to Use & Getting Started

Unlock this app with 5 Redit to access verified runbooks and usage workflows.

Ready to run AGBrowse?

100% Free & Open-Source native desktop software. Star the repository to support development.

More in ai-automation

View all ai-automation apps →
CodexClaw
4.8

CodexClaw

Development discipline + multi-model subagents for OpenAI Codex — 13 dev skills, PABCD workflow, repo map, recall, skill search. Single plugin, no external harness.

Aside Codemode
4.8

Aside Codemode

One call where Aside used to spend fifty — rg-backed search, Aside-shaped file tools, and batched browser work inside a single sandboxed JavaScript block. On a real folder, 55s of find+grep became ...

CLI JAW
4.9

CLI JAW

2-line install personal AI assistant. 11 AI runtime surfaces (Claude, Codex, Cursor, Gemini, Grok, Antigravity, OpenCode, Copilot), 200+ skills, zero ban risk.