guide · ai assistants · mcp

Let an AI assistant compare folders, archives and files with Diffoniq

“What changed between these two releases?” “Is the backup on the NAS complete?” “Which lines differ between these two configs?” An AI assistant that has to answer this on its own reads file after file into its context window, and large folders or large files do not fit. With Diffoniq as an MCP server it asks once and gets the result back as structured data — a summary first, then only what differs; folders, ZIP and 7z archives included. One entry in the assistant’s configuration is all it takes.

What the assistant gets

DiffoniqC.exe /mcp speaks the Model Context Protocol (MCP), the way Claude Code, Claude Desktop, Cursor, GitHub Copilot in VS Code and other assistants connect to tools. It offers three tools. All three only read. None of them changes, moves or deletes a file.

ToolWhat it answers
compare_treeTwo folders or archives: what is different, what exists on one side only, what could not be verified. A folder against a ZIP works too, and archives inside archives are looked into. On request it decides by file content.
compare_fileTwo text files: the lines that differ, with their line numbers on both sides. Case, spaces and trailing blanks can be ignored.
sync_previewWhat a synchronization would do: the folders it would create and the files it would copy, rename or, in mirror mode, delete. A plan, not an action.

Every answer starts with a summary of the whole comparison and then lists the entries, kept short enough for the assistant to read in one go. On a large tree it first names the folders with the most differences; the assistant then looks into one of them, asks only for certain states, or pages on. The tools are also available during the trial period.

Why a tool, and not the assistant on its own

Without a comparison tool, an assistant has to build the answer itself: list both trees, read or hash file after file, and reason about what the lists mean. Everything it reads ends up in its context window, and the window is finite. For two configs that works; for two release folders, a backup on a NAS or a pair of large log files it gets slow, expensive or impossible — and the usual shortcuts have gaps of their own.

The taskThe assistant on its ownWith Diffoniq as its tool
Two large folders Lists and hashes both trees in many steps; the output of every step goes through its context. A tree with thousands of files does not fit. One call. The files never enter the assistant’s context; it reads a summary and a bounded list, and zooms into the folders that matter.
Two large files Reads both files into context, or runs a diff and reads its whole output. Gets only the lines that differ, with line numbers on both sides. The files can be far larger than the assistant could read.
Is the copy really identical? Compares names and sizes, or hashes every file itself. Compares by content with Diffoniq’s hash cache; what it could not verify is reported as unverified, never as identical.
ZIP and 7z archives Extracts them to a temporary folder first — or skips them. Opens archives like folders, also nested, also against a plain folder. Nothing is extracted.
Moved or renamed files Sees one file missing and one file extra. Sees one pair, and reports whether the content changed as well.
Files a cloud service keeps online only Reading a OneDrive or Dropbox placeholder downloads it, silently. Leaves them unread and says how many, unless you allow the download.
“Make them the same” Writes its own copy or delete commands and runs them. Can only show the plan. Copying and deleting stay with you.
Same question tomorrow, other machine Depends on which tools are installed there and how it scripted them. Same factory settings everywhere; every answer lists the options that were in effect.

In short: the assistant asks the question, Diffoniq does the reading. What the assistant gets back is the result, not the data.

1. Tell the assistant where Diffoniq is

The entries below use the installer’s path. If you use the portable ZIP, put in the folder you unpacked it to.

Claude Code — in a terminal:

claude mcp add diffoniq -- "C:\Program Files\Diffoniq\DiffoniqC.exe" --mcp

Claude Desktop — in claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "diffoniq": {
      "command": "C:\\Program Files\\Diffoniq\\DiffoniqC.exe",
      "args": ["/mcp"]
    }
  }
}

Cursor — the same mcpServers block in the project’s .cursor\mcp.json, or in %USERPROFILE%\.cursor\mcp.json for all projects.

VS Code with GitHub Copilot — in the project’s .vscode\mcp.json:

{
  "servers": {
    "diffoniq": {
      "type": "stdio",
      "command": "C:\\Program Files\\Diffoniq\\DiffoniqC.exe",
      "args": ["/mcp"]
    }
  }
}

/mcp, -mcp and --mcp mean the same. In Git Bash use --mcp, because that shell rewrites a leading slash into a path.

2. Ask in plain words

Restart the assistant if it asks you to, then ask as you would ask a colleague:

Most assistants ask you before they use a tool for the first time. Give paths in full (C:\… or \\server\share\…): the server refuses a relative path, because it would depend on the folder the assistant happens to run in.

What it will not do

Good to know

The full reference, including every parameter, is in the help: Scripting and AI Assistants.

Try it with your own folders

Diffoniq runs on Windows 10 and 11, 64-bit. Every feature is available for 30 days before you decide, the MCP server included.

Claude, Cursor, Visual Studio Code, GitHub Copilot, OneDrive, Dropbox and Google Drive are trademarks of their respective owners. Diffoniq is not affiliated with any of them.