Fabric-MCP is a game-changer! Mind-blowing!

[ksylvan]

The Fabric-MCP project is now available. You can get it via PyPi or using the repository.

The project is here: https://github.com/ksylvan/fabric-mcp

With Fabric-MCP at your fingertips, you are only limited by your imagination! This is truly mind-blowing!

I'll write up some documentation about how to connect it to Claude Desktop or other MCP clients.

Demo in Claude Desktop

Here I am showing it connected to Claude Desktop:

Using the Perplexity Integration

And using the Perplexity integration to find the top 5 news headlines with citations:

Here is the second screenshot showing the citations:

Asking Claude to Present a Visual Display of Fabric Patterns

This was in response to: "Get all the available patterns that Fabric-MCP can run and present them in a compact but organized visually interactive way."

Complex Multi-Pattern Workflow

---

1. Use Perplexity's "sonar-pro" model and run the raw_query pattern with the prompt "Give a comprehensive overview of Quantum Computing discoveries over the last year"

2. Use that as input to the "write_essay" pattern, using vendor "Anthropic" and the "claude-sonnet-4-20250514" model.

3. Display the result in a beautiful format.

---

This was the artifact produced: https://claude.ai/public/artifacts/e0d28557-eb8b-48a2-9db2-b06912eeaeb8

[limcheekin]

I just find out you created a better Fabric MCP server than I did, but I'm more focus on transform REST APIs to MCP server without writing single line of code.

Anyone interested can find out more at https://limcheekin.medium.com/building-your-first-no-code-mcp-server-the-fabric-integration-story-90da58cdbe1f

[ksylvan]

@limcheekin I love your approach and I wanted a more full featured MCP server.

I ended up fixing some fabric rest api bugs along the way too.

I'll probably end up reworking a GoLang version of fabric-MCP soon.

[limcheekin]

Look forward to that, I'm definitely an user of your fabric-MCP.

[ksylvan]

By the way you can run fabric --serve on some powerful machine in your LAN (maybe with DeepSeek or other models running locally) and then you can point fabric-MCP to connect to that endpoint.

Now your powerful local LLMs enhanced by Fabric patterns are available for your whole network.

[limcheekin]

Sorry, I don't get it. How do I point fabric-MCP to connect to local endpoint? What endpoint means here Ollama endpoint, OpenAI-compatible endpoint or other type of endpoint?

[ksylvan]

I'll show you an example:

SHAKTI is my WIndows 11 box (with a 4090 installed):

In one terminal, I ssh to the Windows server and run fabric --serve

Then you can define your MCP definition to use the FABRIC_BASE_URL environment variable. For example, in Warp terminal, I can do this:

{
  "FabricMCP": {
    "command": "uv",
    "args": [
      "run",
      "fabric-mcp",
      "--transport",
      "stdio"
    ],
    "env": {
      "FABRIC_BASE_URL": "http://shakti:8080"
    },
    "working_directory": "/Users/kayvan/src/fabric-mcp",
    "start_on_launch": true
  }
}

For Claude Code, I put this into a wrapper shell script and set it up to run in ~/Library/Application Support/Claude/claude_desktop_config.json

#!/bin/bash

# Set up environment
export PATH="/opt/homebrew/bin:/usr/local/bin:$PATH"

# Change to the project directory
cd /Users/kayvan/src/fabric-mcp

# Uncomment this to connect to shakti:8080
# Otherwise it tries to connect to localhost:8080
# export FABRIC_BASE_URL=http://shakti:8080

# Run the fabric-mcp server
exec uv run fabric-mcp --transport stdio

Then I can even ask:

Or ask about local models:

What models are available from Ollama or LM Studio from Fabric?

[ksylvan]

Sorry, I don't get it. How do I point fabric-MCP to connect to local endpoint? What endpoint means here Ollama endpoint, OpenAI-compatible endpoint or other type of endpoint?

If you want a local endpoint, just do:

In one terminal:

fabric --serve

Then set up your tool to run your fabric-mcp similarly to what I show above. You don't need to specify FABRIC_BASE_URL if it is the default (localhost:8080).

[limcheekin]

Wow! Thanks for the detail information.

So, is that mean Claude Desktop can use any LLM models setup/configured in the Fabric?

[ksylvan]

Wow! Thanks for the detail information.

So, is that mean Claude Desktop can use any LLM models setup/configured in the Fabric?

Yes!

[adapoet]

Brilliant! The raw_query tool looks great.