PhotoSweat

MCP

Background removal inside your agent

A remote Model Context Protocol server. One endpoint URL, and Claude, Claude Code, Cursor or VS Code can cut the background out of a photo the user hands it. Nothing to install — no binary, no npm package, no API key, no local model.

POST https://photosweat.shop/mcp · Streamable HTTP · no authentication required.

Listed in the official MCP Registry as shop.photosweat/remove-background — same endpoint, nothing extra to configure.

Add it to your host

Claude — web and desktop

Settings → Connectors → Add custom connector, and paste the URL:

https://photosweat.shop/mcp

Claude Code

claude mcp add --transport http photosweat https://photosweat.shop/mcp

Cursor

Add this to ~/.cursor/mcp.json (or a project-level .cursor/mcp.json):

{
  "mcpServers": {
    "photosweat": { "url": "https://photosweat.shop/mcp" }
  }
}

VS Code and other JSON-configured hosts

In .vscode/mcp.json:

{
  "servers": {
    "photosweat": { "type": "http", "url": "https://photosweat.shop/mcp" }
  }
}

The server speaks Streamable HTTP, so anything that can point at a remote MCP URL works the same way. There is no OAuth dance: the tool is available anonymously.

The one tool: remove_background

A photo in, a transparent PNG out — at the input's pixel size.

ArgumentWhat it is
image_base64A JPEG, PNG or WebP encoded as base64, with or without a data: prefix. Supply this or image_url.
image_urlA publicly reachable image URL. Supply this or image_base64.
backgroundtransparent (default), white or a solid colour.
colorThe solid colour, when background is set to one.
formatOutput format — PNG when transparency is wanted, JPEG when it is not.

What a host sees in tools/list

{
  "name": "remove_background",
  "title": "Remove background",
  "annotations": { "title": "Remove background", "readOnlyHint": false,
                   "destructiveHint": false, "openWorldHint": false },
  "inputSchema": {
    "type": "object",
    "properties": { "image_base64": { … }, "image_url": { … }, … },
    "anyOf": [ { "required": ["image_base64"] }, { "required": ["image_url"] } ]
  }
}

The result is an image content block carrying the PNG, plus a text line saying how many images are left today. When the host asks for structured output the result also carries job_id, bytes, mimeType and remaining_today.

Quota and price

The quota is counted per caller, not per conversation, so an agent that calls the tool in a loop will run out. The result text always says what is left, which lets the host decide whether to keep going.

Questions

Do I need an API key?

No. The endpoint answers anonymous callers on 10 images a day per IP. A free key raises that to 30 a day, and prepaid credits cover anything beyond it.

Does this generate images?

No. It removes the background from a photo you provide and returns a transparent PNG. It does not generate, extend or restore images.

What does the tool return?

An image content block carrying the PNG, plus a text line stating how many images are left today. A structured output also carries job_id, byte size, MIME type and the remaining quota.

What inputs does it accept?

image_base64 or image_url, JPEG/PNG/WebP, up to 10 MB. A long edge over 4096 px comes back scaled down proportionally.

What if the queue is busy?

The tool waits up to 20 seconds and then returns a status_url instead of the image. Poll that URL — do not resend the call, which would charge the quota again.