Weavetab

What Weavetab does

Weavetab exposes browser and desktop automation to MCP clients. It connects directly to live Chromium tabs and registered Electron applications through Chrome DevTools Protocol (CDP) sockets, allowing an agent to navigate pages, inspect interfaces, enter data, manage tabs, and perform UI actions locally.

Its browser representation is a semantic map rather than a full HTML dump. Interactive elements receive session-scoped references, dialogs are surfaced prominently, and repeated link groups can be compressed. Subsequent reads can return only changed nodes through delta mode, which is useful for large single-page applications and repetitive workflows.

The server also includes visual and low-level capabilities. Agents can inspect screenshots, interact with canvas content, control pointer movement, upload files, read console output, inspect network activity, manage cookies and storage, and capture or restore MHTML snapshots. Electron routing lets the same tool family target configured applications such as Discord, VS Code, or Figma.

How it works

MCP clients communicate with Weavetab over JSON-RPC 2.0. The default launch mode uses stdio; an optional WebSocket mode is available with --ws. Local CDP connections and the optional MCP WebSocket listener bind to 127.0.0.1. WebSocket access uses a shared secret stored under ~/.weavetab-system/ws-secret, together with Origin and Host header checks.

The main interaction cycle is to navigate or select a tab, obtain a map or targeted extraction, and act on the returned references. browser_map supports scoping and queries for complex pages. Delta reads reduce repeated output, while visual mapping can produce numbered screenshot marks for canvas-heavy interfaces; those marks can then be passed to browser_click.

browser_burst combines multiple browser actions into one round trip. It supports native actions and higher-level sequences such as filling a form, navigating and reading, or scrolling and reading. Recent successful actions can also be compiled into a burst macro.

Setup and configuration

Node.js 20 or newer is required. The package can be launched without a global installation:

npx -y @weavetab/mcp

A stdio MCP entry can use the following shape:

{
  "mcpServers": {
    "weavetab": {
      "command": "npx",
      "args": ["-y", "@weavetab/mcp"]
    }
  }
}

The documented client locations include Claude Desktop, Claude Code, Cursor, VS Code Copilot, Windsurf, and Antigravity. Electron applications are registered in ~/.weavetab/config.json under an electron object, where each name maps to the executable path. Tools can then receive an app value to target that process instead of a browser tab.

Secrets can be stored with the Weavetab secrets command and supplied through browser_type_secret. Secret values are resolved by the server and masked in responses, snapshots, and audit logs.

Tools and capabilities

• Navigate to URLs or focus configured Electron applications with browser_navigate.
• List, open, switch, and close browser or application tabs with browser_tabs.
• Map interactive DOM content with scoped, queried, visual, or delta reads using browser_map.
• Find text, scrape structured page data, inspect console output, and capture screenshots or PDFs.
• Click, type, select, scroll, upload files, press keys, fill forms, and control pointer gestures.
• Use browser_burst for multi-step actions and reduced round trips.
• Handle JavaScript dialogs, cookies, local storage, session storage, and viewport or device emulation.
• Inspect or stub network responses and retrieve request timing, certificate, and failure details.
• Work with canvas content through image retrieval and simulated drawing.
• Detect and solve supported CAPTCHA flows through CDP mouse interaction without a third-party API.

Limitations and notes

Reference IDs are tied to the current session and become invalid after DOM mutations, so agents should remap when the page changes. browser_eval runs JavaScript through CDP but cannot traverse shadow DOM or cross-origin iframes. XPath selectors do not support shadow DOM in scraping; CSS selectors are required there.

Canvas interaction should use browser_canvas rather than ordinary DOM clicks. Text-only models should generally prefer maps or canvas operations over screenshots. Electron automation requires the target application to be registered and available for CDP attachment. The server is designed for local connections and does not describe a cloud relay workflow.