I Added MCP to My LLM Memory Calculator—Ask AI What Fits

Published 18/10/2026

Description

Will that local LLM fit on your GPU? My LLM Memory Calculator now includes an MCP server, so you can ask the question directly from LM Studio or Claude Code instead of manually filling in a web form. Contents: 0:00 - Introduction 0:14 - Description 0:52 - Website 1:20 - LM Studio 5:48 - Claude Code 8:00 - Conclusion In this hands-on walkthrough, I connect both clients to the calculator’s MCP endpoint, explore its two tools, and test whether Poolside Laguna S 2.1 in Q4_K_M quantization will fit entirely on an RTX 5090 with 32 GB of VRAM. The workflow combines two steps: retrieve the model configuration from Hugging Face, then calculate its estimated memory requirements using the selected quantization and available GPU memory. What you’ll see: Adding the LLM Memory Calculator MCP server to LM Studio’s mcp.json Enabling the integration and configuring tool-approval settings Using the model-configuration tool to retrieve Hugging Face metadata Calling the memory-calculation tool with model and hardware parameters Connecting the same HTTP MCP endpoint to Claude Code Testing the same hardware question in both clients Understanding the difference between fitting entirely in VRAM and running with offloading Checking the tool result rather than relying only on the assistant’s final wording In LM Studio, I use a local Qwen3.8-27B model to query the calculator about a much larger model. The test retrieves approximately 117.6 billion parameters for Laguna S and evaluates the requested configuration against 32 GB of VRAM. The calculator provides an estimate, not a guarantee. The assistant can also phrase its conclusion ambiguously: in my Claude Code test, the answer began with “yes” while explaining that the model would not fit entirely on the GPU. That distinction matters when choosing hardware and planning a local LLM deployment. Whether you prefer the visual web calculator or a conversational MCP workflow, you can now explore model-memory