Connect your AI tool

Fabric Desktop by KratosBI is AI-ready, not AI-dependent. It has no AI of its own: no chat, no model, no AI key or subscription. AI tools you already use (Claude Code, Claude Desktop, Codex, Cursor, VS Code with GitHub Copilot, or any MCP client) can connect to it and use the same features you do, with the permissions you give them. Nothing is sent to an AI service by Fabric Desktop by KratosBI; your tool asks it for things, and you can switch all of it off. Everything here is Preview.

Which tools have been tried#

Only the MCP Python SDK client and an HTTP client are tested, automatically on every build. None of the AI tools below has been tried with Fabric Desktop by KratosBI yet: each setup follows that tool's documented MCP configuration format, but is Untested. The compatibility page is kept up to date; if you try one, open an issue with the tool, its version and what happened.

Connect a tool#

1. Open Settings > Integrations and turn on Let AI tools connect. It's off until you do, and your organization's policy can keep it off.
2. Add a tool: pick which one, give it a name, and choose what it may do. Read metadata (names, structure, schemas, checks; no rows) is always included. Read data reads rows from tables, a few at a time. Run code on this PC runs SQL, Python and notebooks, and asks you every time.
3. Copy setup and paste it into your tool as shown below. Your setup has your tool's own id in place of YOUR-CLIENT-ID and your own path in place of C:\Users\you\.... The command's path stays the same when Fabric Desktop by KratosBI updates.
4. Test connection, then ask your tool something like "What lakehouses can you see in Fabric Desktop?".

Fabric Desktop by KratosBI has to be open while you use it: requests that need your approval appear there, over whatever page you're on. Allow once lets that one request through; Deny refuses it. Your tool can't approve anything itself.

Setup for each tool#

Claude Code#

Run this once in a terminal. Status: Untested.

claude mcp add fabric-desktop -- "C:\Users\you\AppData\Local\FabricDesktop\bin\fabric-desktop-mcp.cmd" --client YOUR-CLIENT-ID

Claude Desktop#

Add this to claude_desktop_config.json (Settings > Developer > Edit config), merging it with any servers already there, then restart Claude Desktop. Status: Untested.

{
  "mcpServers": {
    "fabric-desktop": {
      "command": "C:\\Users\\you\\AppData\\Local\\FabricDesktop\\bin\\fabric-desktop-mcp.cmd",
      "args": [
        "--client",
        "YOUR-CLIENT-ID"
      ]
    }
  }
}

OpenAI Codex CLI#

Add this to ~/.codex/config.toml. Status: Untested.

[mcp_servers.fabric-desktop]
command = "C:\\Users\\you\\AppData\\Local\\FabricDesktop\\bin\\fabric-desktop-mcp.cmd"
args = ["--client", "YOUR-CLIENT-ID"]

Cursor#

Add this to .cursor/mcp.json in your project, or ~/.cursor/mcp.json for every project. Status: Untested.

{
  "mcpServers": {
    "fabric-desktop": {
      "command": "C:\\Users\\you\\AppData\\Local\\FabricDesktop\\bin\\fabric-desktop-mcp.cmd",
      "args": [
        "--client",
        "YOUR-CLIENT-ID"
      ]
    }
  }
}

VS Code (GitHub Copilot agent mode)#

Add this to .vscode/mcp.json in your workspace, then start the server from the file and use Copilot in agent mode. Status: Untested.

{
  "servers": {
    "fabric-desktop": {
      "type": "stdio",
      "command": "C:\\Users\\you\\AppData\\Local\\FabricDesktop\\bin\\fabric-desktop-mcp.cmd",
      "args": [
        "--client",
        "YOUR-CLIENT-ID"
      ]
    }
  }
}

Another MCP client#

Any tool that starts MCP servers over stdio. Most use this shape; check your tool's documentation for where it goes. Status: Untested for any specific tool.

{
  "mcpServers": {
    "fabric-desktop": {
      "command": "C:\\Users\\you\\AppData\\Local\\FabricDesktop\\bin\\fabric-desktop-mcp.cmd",
      "args": [
        "--client",
        "YOUR-CLIENT-ID"
      ]
    }
  }
}

An HTTP client (REST API)#

For scripts and tools that call HTTP instead of starting a command. Turn on Local HTTP API under How tools connect; it listens on this PC only (127.0.0.1). Each tool gets its own token, shown once when you add the tool (Edit > New token makes another). Status: Tested with httpx and the MCP Python SDK.

Base URL: http://127.0.0.1:47615/api/v1
OpenAPI: http://127.0.0.1:47615/api/v1/openapi.json
MCP (Streamable HTTP): http://127.0.0.1:47615/mcp
Header: Authorization: Bearer <the token shown once>

What a connected tool can and can't do#

Code a tool runs gets its own process that's ended afterwards; it can read the attached lakehouse but table writes through Fabric Desktop by KratosBI are refused. Approved Python runs as you and could still write files with the storage access it has, so read it before you allow it: the approval shows it in full. Text the tool reads from notebooks, tables and files is treated as data: it can't approve anything. Every request is recorded under Recent activity, and in an audit trail your IT team can collect (%LOCALAPPDATA%\FabricDesktop\integrations\audit).

These limits apply to what your tool does through Fabric Desktop by KratosBI. A tool that can already run commands on your PC (as most coding agents can) can do anything you can, with or without this connection; see security.

If it doesn't work#