DEV Community

Shanewas Ahmed
Shanewas Ahmed

Posted on

Driving Your Real Microsoft Edge Tabs from AI Agents with edge-agent-bridge

#ai #python #automation #webdev

Most browser automation tools built for AI agents start from an inconvenient premise: they assume your task wants a clean, empty browser window.

Tools like Playwright or Puppeteer boot disposable profiles in headless mode. That's fine for running test suites in continuous integration. It falls flat the moment you want an agent to handle daily engineering work, like approving a pull request behind corporate SAML or pulling a report from an internal intranet portal.

Standard headless browsers don't have your session cookies. They trigger Cloudflare turnstiles, get blocked by bot detectors, and can't pass hardware key 2FA prompts.

I built edge-agent-bridge to sidestep this entirely. Instead of fighting bot walls in a sandbox, it drives the browser tabs you already have open: Edge and Chrome both work.

Why Drive the Live Browser?

When an agent controls your open Edge or Chrome tabs:

  • Authenticated sessions stay valid. You log in once via SSO; your agent uses that existing state.
  • Actions dispatch through Chrome DevTools Protocol as real, trusted input events (isTrusted: true), not injected synthetic JavaScript clicks.
  • Corporate VPN tunnels and client certificates work without extra configuration.

Everything runs locally. A small background daemon listens on 127.0.0.1:18999 and coordinates requests between your agent tools and a companion browser extension.

Quick Setup

1. Install the PyPI Package

The client requires Python 3.10+ and has zero third-party dependencies:

pip install edge-agent-bridge
Enter fullscreen mode Exit fullscreen mode

2. Add the Companion Extension

The extension ships on both stores:

If you prefer building from source, run:

edge-bridge extension open
Enter fullscreen mode Exit fullscreen mode

That command displays the extension directory path. From there, open edge://extensions, toggle Developer mode on, and pick "Load unpacked".

Verify the background service anytime:

edge-bridge daemon status
Enter fullscreen mode Exit fullscreen mode

Three Ways to Use It

1. Element-Ref CLI

The CLI uses an accessibility snapshot model. It maps visible interactive nodes to short ref tokens (e1, e2):

# Capture the active tab
edge-bridge snapshot

# Output format:
# - textbox "Search" [e1] value=""
# - button "Submit" [e2]

# Click or type by ref
edge-bridge fill e1 "agentic automation"
edge-bridge click e2
Enter fullscreen mode Exit fullscreen mode

Refs prevent accidental clicks. If the page reloads or DOM nodes shift before your command finishes, the bridge flags the stale ref and stops rather than clicking the wrong element.

2. Python Scripting

For standalone automation scripts, import the Edge client:

from edge_agent_bridge import Edge

with Edge() as browser:
    snap = browser.snapshot()
    browser.fill("e1", "edge-agent-bridge")
    browser.click("e2")
    browser.wait(text="Done")
Enter fullscreen mode Exit fullscreen mode

The context manager locks onto the active tab from your first query. If you click into another desktop window while the script runs, execution won't wander to the wrong tab.

3. MCP Server for Coding Assistants

You can wire Edge control straight into Claude Code, Cursor, Windsurf, or Gemini CLI:

edge-bridge setup
Enter fullscreen mode Exit fullscreen mode

This registers 29 tools (prefixed with edge_) using the Model Context Protocol stdio transport. Your coding assistant can inspect the DOM tree, fill forms, press key shortcuts, and capture targeted element screenshots.

Security and Privacy

Because the extension interacts with your live tabs, the safety boundaries are intentionally conservative:

  • The daemon binds strictly to loopback (127.0.0.1). It inspects the Host header and rejects external origins or DNS rebinding with HTTP 421.
  • Mutating actions require local authorization tokens stored in your user directory.
  • No network requests leave your machine. There are no telemetry pings, no remote trackers, no crash reporters, and no hosted cloud relays.

Check out the project:

Top comments (0)