Nodegrove VRAM: can I run it?

What Nodegrove VRAM: can I run it? does

Nodegrove VRAM: can I run it? is a Developer Tools MCP server for checking local LLM hardware compatibility. It compares open-weight models with GPUs, estimates memory requirements at several quantisations, and identifies models or hardware changes that could make an otherwise unsuccessful configuration work.

The calculations cover model weights, KV cache, runtime overhead, context length, and a single-stream speed ceiling. Results include the assumptions used, plus links to relevant model and GPU pages on nodegrove.io. The tools are read-only and do not modify models, hardware, or external accounts.

How it works

Model and GPU names can be supplied in common shorthand, by identifier, or by description. Models may be referenced by names such as llama 3.3 70b, model identifiers, or Hugging Face repository IDs. GPUs can be named by product, such as 4090 or M4 Max, or described using memory and bandwidth.

For supported model data, the server reads configuration details and applies fixed formulas. For a Hugging Face repository, it fetches the public config.json and metadata. Its memory model includes weights, KV cache, and overhead; it also accounts for sliding-window layers, hybrid attention, and latent attention where the model configuration exposes those patterns. Fit is assessed against 95% of addressable memory, with results above 85% marked as tight. The speed value is a calculated ceiling rather than a benchmark.

Setup and configuration

The remote endpoint requires no installation:

https://mcp.nodegrove.io/mcp

Add that URL to an HTTP-capable MCP client, or use the client-specific setup documented by the project for Cursor, VS Code, Claude, Codex, Gemini CLI, and other clients. A generic configuration uses the endpoint as the server URL.

For local stdio use, install Node.js 20 or newer and configure the server to run through npx:

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

A self-hosted remote copy can be deployed to Cloudflare using the repository's Worker entry point and Wrangler configuration. The project also exposes its underlying calculation library as @nodegrove/llm-math.

Tools and capabilities

• can_i_run checks a model, GPU, quantisation, and context request, then reports fit status, memory allocation, maximum fitting context, and changes that could resolve a failure.
• what_fits compares the checked models for one card and selects an everyday recommendation, the largest fitting model, the best Q8 option, and the first model that does not fit.
• estimate_vram breaks down memory by quantisation and identifies the smallest common card for each result.
• estimate_from_hf_repo reads a Hugging Face repository configuration and estimates attention details, per-1,000-token context cost, and memory across quantisations.
• list_models and list_gpus return the available model and card records with specifications, IDs, and page links.

Limitations and notes

Nodegrove VRAM: can I run it? does not measure runtime performance. Its speed ceiling is derived from memory bandwidth and active weight reads, so actual inference is expected to be lower. Hardware fit is also an estimate based on published specifications and the server's formulas, not a guarantee for every runtime.

Apple silicon is treated as making approximately 75% of unified memory available to the GPU. The remote service is stateless and keeps no request records, although Cloudflare retains standard edge logs briefly and applies a limit of 120 requests per minute per address. Hugging Face queries retrieve public repository information; the local version contacts Hugging Face only for such requests.

The source code is MIT licensed. The model and GPU data is separately licensed under CC BY 4.0. The remote service is free and does not require an account or API key.