mcp-turboquant

The first MCP server for LLM quantization. Compress any Hugging Face model to GGUF, GPTQ, or AWQ format in a single tool call.

Built on TurboQuant — the unified CLI for model compression.

LLM quantization is one of the most common tasks in the open-source AI workflow, yet there has been no way for AI assistants to do it autonomously. Until now.

With mcp-turboquant, Claude (or any MCP-compatible agent) can:

• Quantize models — convert any HF model to GGUF/GPTQ/AWQ with specified bit widths
• Inspect models — get parameter counts, architecture details, and size estimates
• Recommend settings — analyze your hardware and suggest optimal format + bits
• Check backends — verify which quantization engines are installed

pip install turboquant

Add to your Claude Code MCP settings (~/.claude/settings.json):

{
  "mcpServers": {
    "turboquant": {
      "command": "npx",
      "args": ["-y", "mcp-turboquant"]
    }
  }
}

Or run locally:

{
  "mcpServers": {
    "turboquant": {
      "command": "node",
      "args": ["/path/to/mcp-turboquant/index.js"]
    }
  }
}

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "turboquant": {
      "command": "npx",
      "args": ["-y", "mcp-turboquant"]
    }
  }
}

| Tool | Description | |------|-------------| | quantize | Quantize a HF model to GGUF/GPTQ/AWQ. Params: model (required), format (gguf/gptq/awq), bits (2-8), output (path) | | info | Get model info — param count, architecture, size estimates | | recommend | Hardware-aware recommendation for best format and bit width | | check | List available quantization backends on the system |

Once configured, just ask Claude:

> "Quantize meta-llama/Llama-3.1-8B to 4-bit GGUF"

> "What quantization format should