Remove Metadata by Chatting: MCP Setup for Claude, Cursor, VS Code and ChatGPT

MCP, the Model Context Protocol, lets AI assistants call outside tools. Connect ours and “strip the location from the videos in Downloads” becomes a real action on your machine instead of a website visit. This is the whole setup, with prompts, privacy details and troubleshooting.

Before you start

Configure your app

Claude Desktop

Settings, Developer, Edit Config. Paste this with your key and restart:

{
  "mcpServers": {
    "metadatatools": {
      "command": "npx",
      "args": ["-y", "metadatatools-mcp"],
      "env": { "MRT_API_KEY": "mrt_live_..." }
    }
  }
}

If the server won't start on Windows, use cmd as the command and ["/c", "npx", "-y", "metadatatools-mcp"] as the arguments.

Claude Code

claude mcp add --transport stdio --env MRT_API_KEY=mrt_live_... metadatatools -- npx -y metadatatools-mcp

Cursor and VS Code

Cursor takes the same JSON as Claude Desktop in Settings, MCP or ~/.cursor/mcp.json. VS Code uses .vscode/mcp.json and can prompt for the key at startup so it isn't stored in the file. Both are shown on the MCP page.

Prompts that work

Talk as you would to a colleague. A folder means the supported files directly inside it, not sub-folders.

• “Strip the location from the videos in my Downloads folder.”
• “Remove author and dates from report.pdf before I send it.”
• “Clean every RAW file in ~/Photos/shoot and list what they contained.”
• “Clean the PNGs in my ComfyUI output folder, copies into a folder called clean.”

The assistant calls the clean_files tool. Each file goes to MetadataTools.io over HTTPS, and the clean copy is written beside the original as name-clean.ext (with a number if that name exists). Originals are never modified.

Reading the result

Cleaned 3 of 3 files.
✓ /Users/sam/Downloads/trip.mov → /Users/sam/Downloads/trip-clean.mov: removed 6 fields (com.apple.quicktime.location.iso6709, com.apple.quicktime.make, com.apple.quicktime.model, ...). Verified clean.
✓ /Users/sam/Downloads/report.pdf → /Users/sam/Downloads/report-clean.pdf: removed 5 fields (Author, Creator, Producer, CreationDate, ModDate). Verified clean.
✓ /Users/sam/Downloads/IMG_0042.jpg → /Users/sam/Downloads/IMG_0042-clean.jpg: removed 14 fields (Make, Model, GPSLatitude, ...). Verified clean.
497 left of that allowance today.

The server tells the assistant to call a file clean only if the result says verified clean, which means a second scan of the output found nothing identifying. Otherwise the number of remaining fields is shown.

Who sees what

• Your assistant sees names, paths and the report. File contents go straight from your computer to our server.
• Our server cleans photos and PDFs in memory without storing them. Video and RAW sit in temporary space during processing and the original is deleted when it ends.
• Restrict folders with MRT_ALLOWED_DIRS; other paths are refused.
• Revoke access by deleting the key under Dashboard, API keys.

ChatGPT and claude.ai

Chat apps use the hosted server instead. Add https://metadatatools.io/api/mcp as a custom connector and sign in with your account; no key is needed. These apps can't pass a file dropped into the chat to a connector, so the assistant cleans a file at a public link or offers a one-time upload link. For files on your computer, Claude Desktop, Claude Code, Cursor or VS Code works better.

Allowances and limits

Troubleshooting

For a single file the home page is faster. For unattended jobs, bulk metadata cleaning compares browser, API and assistant.