DEV Community: Arunmozhi Varman

Building a Local, GPU-Accelerated Background Remover MCP Server in Python

Arunmozhi Varman — Fri, 05 Jun 2026 06:24:28 +0000

I have drafted a blog post highlighting your newly published project: bg-vanish-mcp. It covers the motivation, tech stack (DirectML + U2NET), implementation details, and how to use it.

Building a Local, GPU-Accelerated Background Remover MCP Server in Python

As AI assistants like Claude and Windsurf become more integrated into our daily workflows, equipping them with local capabilities (via the Model Context Protocol or MCP) changes the game.

Today, we are announcing bg-vanish-mcp—a local-first, GPU-accelerated background removal MCP server that allows your AI assistants to isolate subjects from images instantly, completely offline, and free of charge.

Here is a look at why we built it, how it works, and how you can run it yourself.

Why Local-First & GPU-Accelerated?

While cloud-based background removal APIs exist, they come with caveats:

Cost: Pay-per-image structures quickly add up.
Privacy: Uploading sensitive images to third-party servers isn't always ideal.
Speed & Latency: Sending high-resolution images back and forth over the network is slow.

bg-vanish-mcp solves this by running U2NET models locally using the rembg library. To keep processing times under a second, it leverages DirectML via ONNX Runtime, allowing hardware acceleration across NVIDIA, AMD, and Intel GPUs natively on Windows—no CUDA or complex SDK configuration required.

How It Works: The Tools

The MCP server exposes two core tools to the AI assistant:

remove_background(input_path, output_path, return_base64)
- Purpose: Operates directly on the host file system. The AI passes a local image path, and the server output saves a clean, transparent PNG in the same directory.
remove_background_base64(image_base64)
- Purpose: Allows direct data passing. The AI can send base64-encoded image data in-context, receive the isolated subject as a base64 string, and render it directly in the chat interface.

Under the Hood: The Code

The server is implemented in Python using the official FastMCP SDK. The core execution provider setup selectively activates the GPU:

import os
import rembg
from mcp.server.fastmcp import FastMCP

mcp = FastMCP("bg-vanish-mcp")

# Read target environment configuration
use_gpu = os.environ.get("USE_GPU", "false").lower() in ("true", "1", "yes")

if use_gpu:
    # Try GPU acceleration via CUDA or DirectML, fallback to CPU
    providers = ['CUDAExecutionProvider', 'DmlExecutionProvider', 'CPUExecutionProvider']
else:
    providers = ['CPUExecutionProvider']

# Initialize the ONNX session
session = rembg.new_session(providers=providers)

By requesting the DmlExecutionProvider (DirectML) before falling back to CPU, the server gains access to native DirectX 12 hardware acceleration on Windows.

Get Started in Seconds

We have published bg-vanish-mcp to PyPI so you can run it instantly without configuring local Python environments.

1. Install & Run via `uvx`

Make sure you have uv installed, then add this block to your Claude Desktop configuration (%APPDATA%\Claude\claude_desktop_config.json):

{
  "mcpServers": {
    "bg-vanish-mcp": {
      "command": "uvx",
      "args": ["bg-vanish-mcp"],
      "env": {
        "USE_GPU": "true"
      }
    }
  }
}

2. Run Locally with Pip

If you prefer standard pip:

pip install "bg-vanish-mcp[dml]"

And point your MCP client to:

bg-vanish-mcp

On first run, the server will download the U2NET ONNX model automatically, bind to your GPU, and be ready to start removing backgrounds in milliseconds.

Check Out the Code

The project is fully open source. Check out the repository, submit issues, or contribute new features:

GitHub: https://github.com/AMV0027/bg-vanish-mcp
PyPI: https://pypi.org/project/bg-vanish-mcp/

How I Built Pop-Upz: An Always-on-Top Floating Web Player Using Chrome's Document PiP API

Arunmozhi Varman — Thu, 04 Jun 2026 18:31:44 +0000

Introduction

Have you ever wanted to keep an eye on a chat, check a stock ticker, monitor a build status board, or scroll through Hacker News while working inside other applications?

Typically, we are forced to resize multiple windows or use second monitors. While browsers support Picture-in-Picture (PiP), the traditional PiP API is strictly limited to <video> elements.

To solve this, I built Pop-Upz—a lightweight Chrome extension that detaches any webpage into a fully borderless, floating, always-on-top window.

In this post, I’ll walk through how I utilized the modern Document Picture-in-Picture API and bypassed cross-origin security restrictions to keep any website floating over your desktop.

🛠️ The Core Technology: Document Picture-in-Picture API

Introduced in Chrome 116, the Document Picture-in-Picture API makes it possible to open an always-on-top window that can be populated with arbitrary HTML content.

Unlike video-only PiP, this opens up endless utility. We can inject an <iframe> inside the PiP window to render an entire website.

Here is the basic framework to launch a Document PiP window from a user gesture:

// Request the Picture-in-Picture Window
const pipWindow = await window.documentPictureInPicture.requestWindow({
  width: 800,
  height: 600
});

// Populate the PiP window's document body
pipWindow.document.write(`
  <html>
    <body>
      <iframe src="https://example.com" style="width:100%; height:100%; border:none;"></iframe>
    </body>
  </html>
`);
pipWindow.document.close();

🛑 The Challenge: Bypassing CSP and Frame Restrictions

If you try to load major websites (like YouTube, GitHub, or LinkedIn) inside an iframe, you will immediately run into console errors:

Refused to display 'https://...' in a frame because it set 'X-Frame-Options' to 'sameorigin'.

Websites use headers like X-Frame-Options and Content-Security-Policy (CSP) to prevent clickjacking attacks by forbidding their pages from being embedded in iframes on other domains.

The Solution: Manifest V3 `declarativeNetRequest`

To bypass this limitation for our floating player, we can leverage Chrome's declarative rules to strip these blocking headers only when the iframe is requested inside our extension's subframe context.

Here is the ruleset configuration (rules.json):

[
  {
    "id": 1,
    "priority": 1,
    "action": {
      "type": "modifyHeaders",
      "responseHeaders": [
        { "header": "X-Frame-Options", "operation": "remove" },
        { "header": "Frame-Options", "operation": "remove" },
        { "header": "Content-Security-Policy", "operation": "remove" }
      ]
    },
    "condition": {
      "resourceTypes": ["sub_frame"]
    }
  }
]

By removing these headers dynamically at the browser engine level, we can frame any website seamlessly inside our floating window.

⚡ Solving the Chrome Transient Activation Bug

One of the trickiest parts of building this extension was handling Chrome's User Gesture requirement. The browser prevents scripts from launching pop-ups programmatically unless triggered by an explicit click.

If you have any asynchronous delay (await statements or microtask yields) between the user's click and the requestWindow() call, Chrome immediately invalidates the gesture token, throwing this error:

"Could not launch Picture-in-Picture window. Please interact with the page first."

To solve this, I designed Pop-Upz to run the PiP instantiation completely synchronously inside the button click handler, only running the document generation logic inside the resolving .then() promise block:

// Synchronously bound to the click event to preserve transient gesture
window.documentPictureInPicture.requestWindow({
  width: 800,
  height: 600
})
.then((pipWindow) => {
  // Safe to write document asynchronously here!
  pipWindow.document.write(...);
  pipWindow.document.close();
});

🎨 Clean and Minimalist Design

Instead of adding heavy settings menus or floating control bars that clutter the screen, Pop-Upz detaches into a 100% borderless, full-bleed viewport.

Users can trigger it in two ways:

Clicking the Pop-Upz Toolbar Icon
Right-clicking anywhere on a webpage and selecting "Pop Out with Pop-Upz ✦"

Once detached, you can drag the corners of the native window to resize it to any layout you prefer (Mobile view, Tablet view, Wide, etc.).

🚀 Check Out the Project

The project is fully open-source and ready to load as a developer unpacked extension.

GitHub Repository: AMV0027/Pop-Upz
LinkedIn: Arunmozhi Varman K

Feel free to clone the repo, try it out, and star the project if you find it useful! Let me know in the comments how you plan to use always-on-top floating web pages in your workflow.

DEV Community: Arunmozhi Varman

Building a Local, GPU-Accelerated Background Remover MCP Server in Python

Building a Local, GPU-Accelerated Background Remover MCP Server in Python

Why Local-First & GPU-Accelerated?

How It Works: The Tools

Under the Hood: The Code

Get Started in Seconds

1. Install & Run via uvx

2. Run Locally with Pip

Check Out the Code

How I Built Pop-Upz: An Always-on-Top Floating Web Player Using Chrome's Document PiP API

Introduction

🛠️ The Core Technology: Document Picture-in-Picture API

🛑 The Challenge: Bypassing CSP and Frame Restrictions

The Solution: Manifest V3 declarativeNetRequest

⚡ Solving the Chrome Transient Activation Bug

🎨 Clean and Minimalist Design

🚀 Check Out the Project

1. Install & Run via `uvx`

The Solution: Manifest V3 `declarativeNetRequest`