Tactab 🌐🤖

Tactile Browser Control & Multimodal Vision Bridge for AI Agents

Tactab (Tactile + Tab) is a high-performance, local, and open-source bridge that gives any MCP-compliant AI client (Cursor, Claude Desktop, Antigravity, Windsurf, Cline, etc.) live visual eyes and tactile hands in your Google Chrome browser using the Model Context Protocol (MCP), WebSockets, and a Chrome Extension (Manifest V3).

⚡ Highlights

• 🌐 Works with Any Plan (Free or Paid): Seamlessly connects whether you are on free tiers or paid Pro/Enterprise plans. Zero subscriptions or paid cloud automation platforms (like Browserbase or MultiOn) required.

• 👁️ Visual Multimodal Vision & DOM Control: Enables your AI to capture high-res PNG screenshots for visual layout inspection, alongside full DOM scraping, button clicks, form filling, and navigation.

• 🔒 Private & 100% Local: Operates strictly over local loopback (127.0.0.1). Your session cookies, authenticated tabs (AWS, Jira, GitHub), and local dev servers (localhost:3000) never leave your machine.

• 🧩 Zero Extension Bloat (One Extension for All Agents): Eliminates the need to install separate, heavy browser extensions for each AI tool. Tactab acts as a single, lightweight gateway that connects Claude, Cursor, Antigravity, or custom agents through standard MCP.

• 🔄 Universal MCP Standard: Plug-and-play with Claude Desktop, Cursor, Antigravity, Windsurf, or custom AI agents over standard I/O (stdio).

Related MCP server: real-browser-mcp

🏗️ Architecture

┌─────────────────────────────────┐         STDIO (JSON-RPC)         ┌───────────────────────────────┐
│     AI Agent / IDE Client       │  ◄─────────────────────────────► │    Tactab MCP Server (Node)   │
│ (Cursor, Claude, Antigravity)   │                                  │         tactab/server         │
└─────────────────────────────────┘                                  └───────────────┬───────────────┘
                                                                                     │
                                                                            WebSocket (ws://127.0.0.1:8765)
                                                                                     │
┌─────────────────────────────────┐     chrome.tabs.sendMessage     ┌────────────────▼───────────────┐
│        Webpage DOM (Tab)        │  ◄─────────────────────────────► │     Tactab Extension (MV3)    │
│          (content.js)           │                                  │    (background.js + badge)    │
└─────────────────────────────────┘                                  └───────────────────────────────┘

1. AI Agent invokes the automate_website tool over standard input/output (stdio).

2. Tactab Bridge Server (bridge.js) listens on 127.0.0.1:8765 and translates tool calls into JSON WebSocket messages.

3. Tactab Chrome Extension Service Worker (background.js) receives commands, manages connection status, and routes to active tabs.

4. Content Script (content.js) executes actions inside the active web page and returns structured results back up the pipeline.

📂 Project Structure

tactab/
│
├── extension/                      # Chrome Extension (Manifest V3)
│   ├── manifest.json               # Extension manifest declaration
│   ├── background.js               # Service worker WebSocket client & badge manager
│   └── content.js                  # DOM interaction & scraping execution engine
│
├── server/                         # Local Node.js MCP Server
│   ├── package.json                # Dependencies (@modelcontextprotocol/sdk, ws)
│   └── bridge.js                   # MCP Stdio transport + WebSocket bridge
│
├── AGENTS.md                       # Universal Multi-Agent Handoff Guardrail
├── PLAN.md                         # Active milestone tracker & Decision Log
├── mcp_config.example.json         # Universal MCP configuration template
├── LICENSE                         # MIT License
├── .gitignore                      # Git ignore file (excludes plane.md, node_modules)
└── README.md                       # Documentation & setup guide

🚀 Quick Start Guide

1. Install Server Dependencies

Open your terminal in the server/ directory:

cd server
npm install

(On Windows systems where script execution policies restrict npm, run npm.cmd install)

2. Load the Chrome Extension

1. Open Google Chrome and navigate to chrome://extensions/.

2. Toggle on Developer mode in the top-right corner.

3. Click the Load unpacked button in the top-left corner.

4. Select the extension/ folder in this repository.

5. Tactab — Browser MCP Bridge will now appear in your extensions list.

   • When connected to the local MCP server, the badge displays green ON.

   • When disconnected, it displays red OFF and automatically retries every 5 seconds.

3. Configure Your AI Client

Add Tactab to your AI client's MCP configuration using your absolute path to server/bridge.js:

A. Claude Desktop

Config file location:

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "tactab": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/TO/tactab/server/bridge.js"
      ]
    }
  }
}

Windows Note: Use escaped backslashes in paths, e.g.: "C:\\path\\to\\tactab\\server\\bridge.js"

B. Cursor IDE

Open Cursor Settings ➔ Features ➔ MCP Servers (or edit .cursor/mcp.json):

{
  "mcpServers": {
    "tactab": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/TO/tactab/server/bridge.js"
      ]
    }
  }
}

C. Antigravity / Windsurf / Cline

Add the exact same JSON block into your environment's MCP server configuration file.

🛠️ Supported Browser Actions

The automate_website tool supports the following actions on any standard webpage:

💡 Example Prompts to Ask Your AI

Once configured, simply instruct your AI in natural language:

• 📸 "Take a screenshot of my current tab and inspect the layout design."

• 📊 "Look at the charts on my active tab and summarize the visual data."

• 🔍 "Look at my active Chrome tab and summarize what you see."

• ✍️ "Fill in the login form with test credentials and submit."

• 🔗 "Extract all product titles and links visible on the page."

• 📜 "Scroll down 800 pixels and extract the pricing table."

🤖 Multi-Agent Handoff Protocol (The "Triple-Anchor" Architecture)

Whether you are switching between specialized models (e.g. Cursor for rapid coding, Claude for system architecture), managing token budgets, or navigating provider rate limits, this repository includes the Triple-Anchor Architecture for seamless context transfer with zero loss of progress or hallucination.

┌───────────────────────────────────────────────────────────────┐
│ Anchor 1: AGENTS.md (Universal Agent Guardrail)               │
│ -> Tells any newly opened agent: "Read PLAN.md and Git first" │
└──────────────────────────────┬────────────────────────────────┘
                               │
       ┌───────────────────────┴───────────────────────┐
       ▼                                               ▼
┌───────────────────────────────┐   ┌───────────────────────────────┐
│ Anchor 2: PLAN.md             │   │ Anchor 3: Git Status & Diff   │
│ (The Intent & Architecture)   │   │ (The Ground Truth of Code)    │
│ • Completed tasks             │   │ • Exact lines of code changed │
│ • Next pending tasks          │   │ • Clean syntax state          │
│ • "Decision Log" & Gotchas    │   │ • Zero hallucinated changes   │
└───────────────────────────────┘   └───────────────────────────────┘

How to Use in Any Project

Copy AGENTS.md and PLAN.md into the root of any codebase:

1. Anchor 1 (AGENTS.md): Automatically read by modern AI tools (Cursor, Antigravity, Claude, ChatGPT CLI). It instructs the model to inspect PLAN.md and git status before modifying any files.

2. Anchor 2 (PLAN.md): Contains the live checklist and Decision Log. The Decision Log is critical: it prevents the incoming agent from accidentally reverting intentional architecture choices (such as why a heartbeat ping was added).

3. Anchor 3 (Git Milestones): When your current agent approaches its token limit, tell it:

   "Commit your progress to Git and update PLAN.md."

4. Instant Continuation: Open your next AI agent and simply prompt:

   "Continue." The new model reads AGENTS.md ➔ PLAN.md ➔ git status, and picks up immediately with zero lost context.

🔧 Troubleshooting

• "Chrome extension is disconnected": Make sure Google Chrome is open, the extension is loaded, and you have an active standard website open (not a restricted internal page like chrome://extensions or about:blank).

• Port Customization: By default, the bridge uses WebSocket port 8765 (avoiding conflict with standard HTTP 8080 proxies). To use a different port, set the WS_PORT environment variable before launching (e.g. set WS_PORT=8090 / export WS_PORT=8090) and update the port in extension/background.js.

⚠️ Legal Disclaimer & Responsible Use

IMPORTANT NOTICE: This software is provided for personal workflow automation, educational research, and authorized testing purposes only.

1. Compliance with Terms of Service: Users are solely responsible for ensuring that their automated actions, scraping requests, and web interactions comply with all applicable local, national, and international laws, as well as the Terms of Service, Acceptable Use Policies, and robots.txt guidelines of any websites visited.

2. No Unauthorized Access: This software must not be used to bypass authentication barriers, paywalls, CAPTCHAs, rate limits, or security controls, nor to access or extract proprietary, copyrighted, or sensitive personal data without explicit permission.

3. Limitation of Liability: The author(s) and contributor(s) of this project assume no liability or responsibility for any misuse, website bans, legal disputes, data loss, damages, or consequences resulting from the installation or execution of this software. By using this project, you agree to assume all associated risks and responsibilities.

📄 License

Released under the MIT License. Free for open source and commercial use.