@nodegrove/vram-mcp

@nodegrove/vram-mcp

An MCP server that tells your AI assistant whether an open-weight LLM fits your GPU, and what would fit instead. The arithmetic behind nodegrove.io, for any model on Hugging Face and any card. Free, read-only, no account or key.

Can my RTX 4090 run Llama 3.3 70B?

No: Llama 3.3 70B at Q4_K_M with 8,192 tokens of context needs 45.8 GB, and the RTX 4090 holds 22.8 GB after headroom. Short by 23 GB. The RTX 6000 Ada (48 GB) is within a whisker: with the KV cache at Q8 it needs 44.4 GB and fits.

Use it

Over stdio, with Node.js 20 or newer:

{
  "mcpServers": {
    "nodegrove-vram": {
      "command": "npx",
      "args": ["-y", "@nodegrove/vram-mcp"]
    }
  }
}

Or remotely, with nothing to install: https://mcp.nodegrove.io/mcp (Streamable HTTP, no authentication). Setup for Claude, Cursor, VS Code, ChatGPT and other clients: nodegrove.io/mcp.

Tools

Tool	Answers
can_i_run	Fits, tight or no, the memory split, a speed ceiling, the longest context that fits, and on a no every change that would make it fit
what_fits	Every model checked on one card, with a recommended pick, the largest that fits and the best at Q8
estimate_vram	Weights, KV cache and overhead at each quantisation, and the smallest common card that holds each
estimate_from_hf_repo	Any Hugging Face repo, read from its config.json, with memory at every quantisation
list_models, list_gpus	The models and cards, with their specs and pages

Every figure is a stated formula over config.json values and the makers' specs, never a benchmark; speeds are ceilings. The method is at nodegrove.io/data.

Privacy

The local server sends nothing to Nodegrove. It contacts huggingface.co, and only when you ask about a repo there, to read that repo's public config.json.

Licence

MIT for the code; the bundled model and GPU data is CC BY 4.0. Credit Nodegrove (nodegrove.io).