mcp-gpu-server
mcp-name: io.github.moezdil/mcp-gpu-server
# mcp-gpu-server
[](https://pypi.org/project/gpu-server/)
An MCP server that exposes NVIDIA GPU metrics as tools. Once connected, any MCP-compatible client can query your GPU status in real time directly from a conversation.
## What it does
Instead of running nvidia-smi manually, you ask your AI assistant and it calls these tools automatically:
gpu_info GPU name, driver version, CUDA version
gpu_utilization core utilization % and memory controller busy %
gpu_vram total, used, free VRAM in MiB and usage %
gpu_temperature GPU core temperature in Celsius
gpu_stats everything above in one call
Example response from gpu_stats:
{
"count": 1,
"gpus": [{
"index": 0,
"name": "NVIDIA L40S",
"driver": "580.126.09",
"cuda": "13.0",
"temp_c": 29,
"gpu_pct": 0,
"mem_pct": 0,
"vram": {
"total_mib": 46068,
"used_mib": 610,
"free_mib": 45457,
"pct": 1.3
}
}]
}
## How it works
Queries NVML (pynvml) directly when available. Falls back to nvidia-smi subprocess if NVML is not accessible. Returns clean JSON in both cases. The nvidia-smi query does not include the CUDA version, so `cuda` is `null` on that path.
## Install
pip install gpu-server
The package moved from `mcp-gpu-server` to `gpu-server` on PyPI in 0.2.1. The old name is frozen at 0.1.3, which does not start with mcp 2.x. The `mcp-gpu-server` command still works, and since 0.2.2 `gpu-server` does too.
Listed on the official [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.moezdil/mcp-gpu-server`.
## Connect to your MCP client
Add this to your MCP client config file:
{
"mcpServers": {
"gpu": {
"command": "mcp-gpu-server"
}
}
}
Or without installing anything, using [uv](https://docs.astral.sh/uv/):
{
"mcpServers": {
"gpu": {
"command": "uvx",
"args": ["gpu-server"]
}
}
}
## Run tests
uv run pytest tests/test_server.py # no GPU needed
uv run python tests/test_gpu.py # on a GPU host
## Tested on real hardware
Release 0.2.2 was installed from PyPI into a clean virtual environment on a Linux host and tested against a real GPU on 2026-10-21.
| | |
|---|---|
| GPU | NVIDIA GeForce RTX 3090, 24 GB |
| Driver / CUDA | 580.178.04 / 13.0 |
| Python | 3.14.4 |
| Packages | gpu-server 0.2.2, mcp 2.3.0, nvidia-ml-py 13.615.71 |
Install the published package and run the GPU integration test:
python3 -m venv venv
venv/bin/pip install gpu-server==0.2.2
git clone --depth 1 https://github.com/moezdil/mcp-gpu-server.git repo
venv/bin/python repo/tests/test_gpu.py
Result: all 37 checks passed. They exercise the NVML path; the nvidia-smi fallback is only checked for presence, not parsed.
Then call every tool through a real MCP client over stdio:
import anyio, json, sys
from mcp import Client, StdioServerParameters
async def main():
async with Client(StdioServerParameters(command=sys.argv[1])) as c:
print("protocol:", c.protocol_version)
for t in (await c.list_tools()).tools:
r = await c.call_tool(t.name, {})
print(t.name, t.annotations.read_only_hint, json.dumps(r.structured_content))
anyio.run(main)
Saved as `e2e.py` and run with `venv/bin/python e2e.py venv/bin/gpu-server`. Result: the client and server negotiated protocol `2026-07-28`, all five tools are listed as read-only, and every call returned structured content without errors:
gpu_stats {"count": 1, "gpus": [{"index": 0, "name": "NVIDIA GeForce RTX 3090",
"driver": "580.178.04", "cuda": "13.0", "temp_c": 42, "gpu_pct": 0, "mem_pct": 0,
"vram": {"total_mib": 24576, "used_mib": 451, "free_mib": 24124, "pct": 1.8}}]}
## Requirements
Python 3.10 or higher, MCP Python SDK 2.x. NVIDIA GPU with drivers installed on the host machine.
## Links
- PyPI: https://pypi.org/project/gpu-server/
- MCP Registry entry: https://registry.modelcontextprotocol.io/v0/servers?search=io.github.moezdil/mcp-gpu-server
- Model Context Protocol: https://modelcontextprotocol.io
- MCP Python SDK: https://github.com/modelcontextprotocol/python-sdk
- uv: https://docs.astral.sh/uv/
- NVML (nvidia-ml-py): https://pypi.org/project/nvidia-ml-py/
TDQS
Scored across 5 tools
Tools are mostly distinct, but 'gpu_info' and 'gpu_stats' overlap somewhat as stats includes info content. However, descriptions clarify the difference (basic vs. full snapshot). The other three tools are clearly separate.
All tool names follow a consistent 'gpu_<noun>' pattern in snake_case. The nouns (info, stats, temperature, utilization, vram) are descriptive and predictable.
Five tools is an appropriate scope for a GPU monitoring server. Each tool covers a distinct aspect without unnecessary redundancy, and the count is not overwhelming.
The tool set covers the main GPU monitoring needs: basic info, comprehensive stats, temperature, utilization, and VRAM. Minor gaps like power usage or process listing are absent but not critical for common use cases.