Fused MCP Agents

Fused MCP Agents: Setting up MCP Servers for Data

Documentation   🌪️    Read the announcement    🔥   

Join Discord

MCP servers allow LLMs like Claude to make HTTP requests, connecting them to APIs & executable code. We built this repo for ourselves & anyone working with data to easily pass any Python code directly to your own desktop Claude app.

This repo offers a simple step-by-step notebook workflow to setup MCP Servers with Claude's Desktop App, all in Python built on top of Fused User Defined Functions (UDFs).

Requirements

• Python 3.11
• Latest Claude Desktop app installed (macOS & Windows)

If you're on Linux, the desktop app isn't available so we've made a simple client you can use to have it running locally too!

You do not need a Fused account to do any of this! All of this will be running on your local machine.

Installation

• Clone this repo in any local directory, and navigate to the repo:

  git clone https://github.com/fusedio/fused-mcp.git
  cd fused-mcp/

• Install uv if you don't have it:

  macOS / Linux:

  curl -LsSf https://astral.sh/uv/install.sh | sh

  Windows:

  powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
  

• Test out the client by asking for its info:

  uv run main.py -h

• Start by following our getting-started notebook fused_mcp_agents.ipynb in your favorite local IDE to get set up and then make your way to the more advanced notebook to make your own Agents & functions

!Notebook

Repository structure

This repo is build on top of MCP Server & Fused UDFs which are Python functions that can be run from anywhere.

Support & Community

Feel free to join our Discord server if you want some help getting unblocked!

Here are a few common steps to debug the setup:

• Running uv run main.py -h should return something like this:

!uv helper output function

• You might need to pass global paths to some functions to the Claude_Desktop_Config.json. For example, by default we only pass uv:

{
    "mcpServers": {
        "qgis": {
            "command": "uv",
            "args": ["..."]
        }

    }
}

But you might need to pass the full path to uv, which you can simply pass to common.generate_local_mcp_config in the notebook:

# in fused_mcp_agents.ipynb
import shutil 

common.generate_local_mcp_config(
    config_path=PATH_TO_CLAUDE_CONFIG,
    agents_list = ["get_current_time"],
    repo_path= WORKING_DIR,
    uv_path=shutil.which('uv'),
)

Which would create a config like this:

{
    "mcpServers": {
        "qgis": {
            "command": "/Users//.local/bin/uv",
            "args": ["..."]
        }

    }
}

• If Claude runs without showing any connected tools, take a look at the MCP Docs for troubleshooting the Claude Desktop setup

Contribute

Feel free to open PRs to add your own UDFs to udfs/ so others can play around with them locally too!

Using a local Claude client (without Claude Desktop app)

If you are unable to install the Claude Desktop app (e.g., on Linux), we provide a small example local client interface to use Claude with the MCP server configured in this repo:

NOTE: You'll need an API key for Claude here as you won't use the Desktop App

• Create an Anthropic Console Account

• Create an Anthropic API Key

• Create a .env:

  touch .env

• Add your key as ANTHROPIC_API_KEY inside the .env:

  # .env
  ANTHROPIC_API_KEY = "your-key-here"
  

• Start the MCP server:

  uv run main.py --agent get_current_time

• In another terminal session, start the local client, pointing to the address of the server:

  uv run client.py http://localhost:8080/sse