Agent Webbridge

What agent-webbridge does

The agent-webbridge MCP server lets an MCP client operate a real Google Chrome installation rather than a separate headless or hosted browser. It uses the browser's existing profiles and login state, so an agent can work with pages already authenticated in Chrome. Operation is local: commands are sent to a router on 127.0.0.1, with no cloud service required by the described architecture.

The server exposes browser controls for opening and locating tabs, reading page structure, running page JavaScript, clicking controls, filling standard and rich-text fields, and uploading files. It can also save screenshots and PDFs, inspect network requests, list tabs and tab groups, and close tabs or complete sessions.

Profiles identify which Chrome account to drive. Sessions group task tabs into a Chrome tab group. This supports workflows that use several accounts or need multiple tabs operating at the same time.

How it works

An MCP request reaches a local HTTP router through a command endpoint. The router selects the daemon associated with the requested profile, and that daemon communicates with the Agent WebBridge Chrome extension over WebSocket. The extension uses Chrome DevTools Protocol through chrome.debugger.

Each tab receives its own debugger attachment. As a result, several tabs within a profile can be handled concurrently instead of sharing one current-tab connection. Multiple profiles are supported through separate daemons, with the router directing requests by profile name. The tabId and profile values allow callers to select the execution target where supported by the client interface.

Setup and configuration

Install the package globally with Node.js, then run awb setup for each Chrome profile. The setup flow opens the Agent WebBridge Chrome Web Store listing and waits for the extension to be added. Profiles can later be started with awb up and stopped with awb down.

The documented requirements are macOS or Windows, Google Chrome, and Node.js 18 or newer. Add the MCP server using the command npx -y agent-webbridge mcp; the README provides configuration examples for Claude Desktop, Claude Code, Cursor, and Windsurf. The local command router listens on 127.0.0.1:10086.

Tools and capabilities

The agent-webbridge MCP server provides these native operations:

• Navigate to a URL, find tabs by title or URL, and list or close tabs.
• Produce accessibility snapshots with stable @e references for later interaction.
• Perform synthetic DOM clicks or CDP mouse clicks marked as trusted for pages that reject ordinary clicks.
• Fill native inputs and contenteditable fields.
• Assign files to file inputs and intercept native file-picker dialogs during uploads.
• Evaluate JavaScript in the page.
• Save full-page or element screenshots and export pages as PDFs.
• Start, stop, list, and inspect captured network requests.
• Close a session and its associated tab group.

The repository also documents Node.js and zero-dependency Python SDK examples, in addition to MCP access.

Limitations and notes

The browser must be Google Chrome on macOS or Windows, and the extension must be installed and connected for the selected profile. The material does not describe Linux support. Local execution keeps commands on the stated loopback address, but browser actions still affect real accounts and pages; use the tool only where automation is authorized and follow each site's terms.

Existing login state does not guarantee that every site permits automation. The README notes that sites can still limit activity from a user's Chrome profile. Benchmark timings vary by machine and network, so the documented parallel-tab comparison should not be treated as a universal performance guarantee.