toolbox-mcp

An MCP server that gives AI agents real tools for the things it genuinely cannot do on its own. Read exact text from images and scans (OCR), process audio and video files, read and edit PDFs, work with images and Excel spreadsheets, run exact math, and more. Everything runs on your own machine over stdio. No network calls, nothing leaves your computer.

Inspired by the browser toolbox at footrue.com, reimplemented in Node so agents like Claude can call the same kind of tools directly.

Why

Modern models reason, see and code well, but there are whole categories they still cannot handle: they cannot open an audio or video file, they misread exact characters in an image, and they quietly get deterministic tasks (arithmetic, timezone math, regex) subtly wrong. This server hands those jobs to real, tested functions, so the agent gets the exact right answer, without burning tokens and without sending your data anywhere.

Two capabilities are worth calling out because a good model genuinely cannot do them at all:

• Audio and video — an LLM cannot open a media file. These tools inspect, convert, trim, and pull frames or audio from video, using a bundled ffmpeg (no install, no network).
• OCR — vision models describe images well but misread exact digits, serial numbers and table cells. ocr_image reads the real text with a local, offline model.

Tools

Reading images and media

PDF

Files and office

Image

Compute (the things LLMs get wrong)

Crypto and security

Data and text

PDF and image tools read and write files by path, so the agent can hand you a document and get a processed file back. Everything runs on your own machine.

Setup

toolbox-mcp is a standard stdio MCP server, so it works in any MCP client. The easiest way is npx, which downloads and runs it with no install or build step. Just add one of the configs below.

Claude Desktop / Cursor

Edit claude_desktop_config.json (Claude) or the MCP settings (Cursor):

{
  "mcpServers": {
    "toolbox": {
      "command": "npx",
      "args": ["-y", "toolbox-mcp"]
    }
  }
}

VS Code (Copilot agent mode)

Create .vscode/mcp.json in your workspace, or run MCP: Open User Configuration from the Command Palette. Note that VS Code uses the servers key:

{
  "servers": {
    "toolbox": {
      "command": "npx",
      "args": ["-y", "toolbox-mcp"]
    }
  }
}

Google Antigravity

Open Settings > Customizations > Open MCP Config (this edits ~/.gemini/config/mcp_config.json), then add:

{
  "mcpServers": {
    "toolbox": {
      "command": "npx",
      "args": ["-y", "toolbox-mcp"]
    }
  }
}

Save and hit refresh in the Installed MCP Servers section.

After configuring, restart or refresh the client and the tools appear automatically. Then just ask, for example, "read this PDF", "convert this image to webp", or "what is 90 km/h in m/s", and the agent will call the tool.

Run from source

To hack on it locally instead of using npx:

git clone https://github.com/medoxisto/toolbox-mcp
cd toolbox-mcp
npm install
npm run build   # or: npm run dev

Then point your client at node /absolute/path/to/toolbox-mcp/dist/index.js.

Privacy

Everything runs locally over stdio. There are no network requests and no telemetry. The compute, data and crypto tools are pure functions; the PDF and image tools read and write only the file paths you give them. Your data never leaves your machine.

Contributing

The toolbox is easy to extend: each tool is a small object with a name, description, a zod schema, and a handler that returns a string. See src/tools. PRs that add useful, deterministic, network-free tools are welcome.

License

MIT. See LICENSE.