Unsloth MCP Server

An MCP server for Unsloth - a library that makes LLM fine-tuning 2x faster with 80% less memory.

What is Unsloth?

Unsloth is a library that dramatically improves the efficiency of fine-tuning large language models:

• Speed: 2x faster fine-tuning compared to standard methods
• Memory: 80% less VRAM usage, allowing fine-tuning of larger models on consumer GPUs
• Context Length: Up to 13x longer context lengths (e.g., 89K tokens for Llama 3.3 on 80GB GPUs)
• Accuracy: No loss in model quality or performance

Unsloth achieves these improvements through custom CUDA kernels written in OpenAI’s Triton language, optimized backpropagation, and dynamic 4-bit quantization.

Features

• Optimize fine-tuning for Llama, Mistral, Phi, Gemma, and other models
• 4-bit quantization for efficient training
• Extended context length support
• Simple API for model loading, fine-tuning, and inference
• Export to various formats (GGUF, Hugging Face, etc.)

Installation

Current state: install from source (below). The package is prepared for npm (name, bin, server.json registry manifest are in place) but is not yet published — the npx-based config in the next section works only after the npm publish lands.

1. Install Unsloth on the machine that will run the fine-tuning: pip install unsloth

2. Build the server from source:

   git clone https://github.com/OtotaO/unsloth-mcp-server.git
   cd unsloth-mcp-server
   npm install
   npm run build
   

3. Point your MCP client (Claude Desktop, Claude Code, Cline, Cursor, …) at the built entrypoint:

   {
     "mcpServers": {
       "unsloth": {
         "command": "node",
         "args": ["/path/to/unsloth-mcp-server/build/index.js"],
         "env": { "HUGGINGFACE_TOKEN": "your_token_here" }
       }
     }
   }
   

   HUGGINGFACE_TOKEN is optional (only needed for gated/private models). Omit the env block if you don’t use one.

After npm publish (planned)

Once published to npm as unsloth-mcp-server, the same clients can run it over stdio via npx with no clone/build:

{
  "mcpServers": {
    "unsloth": {
      "command": "npx",
      "args": ["-y", "unsloth-mcp-server"],
      "env": { "HUGGINGFACE_TOKEN": "your_token_here" }
    }
  }
}

For Claude Code: claude mcp add unsloth -- npx -y unsloth-mcp-server

Available Tools

check_installation

Verify if Unsloth is properly installed on your system.

Parameters: None

Example:

const result = await use_mcp_tool({
  server_name: "unsloth",
  tool_name: "check_installation",
  arguments: {}
});

list_supported_models