# Liquid Compute GPU test

> Liquid Compute GPU test is an MCP server for cloud and infrastructure. Hosted GPU acceptance tests for NVIDIA GPU nodes you administer, run over SSH via MCP. It runs as a hosted remote server, so there is nothing to install locally.

- **Kind:** MCP server
- **Platforms:** Web, API, CLI, self-hosted
- **Website:** https://liquidcompute.com/gpu-test
- **Repository:** https://github.com/liquid-compute/gcat
- **MCP Registry name:** com.liquidcompute/gcat
- **Categories:** [cloud and infrastructure](https://ai-powered.com/categories/cloud-and-infrastructure)
- **Listing updated:** 2026-11-18

## About

The README says the hosted service connects to your nodes over SSH from Liquid Compute's own servers and runs NVIDIA DCGM diagnostics, a CUTLASS GEMM benchmark, HBM and PCIe bandwidth tests, and NCCL all-reduce, then reports each measurement against thresholds set for the chosen hardware profile.

The workflow is driven from an AI agent. The agent picks a profile matching the nodes, and you add a single authorized_keys line for a root or passwordless-sudo user. Test files are copied over for the run and removed afterwards, and the page states that nothing is installed on the node.

It suits people who run NVIDIA A10, H100, H200, B200 or B300 nodes and want a shareable acceptance report before putting hardware into service. Consumer GeForce RTX cards are not supported, and the GPUs must be idle for the full run, which can take 90 minutes or longer.

## Features

- **Hardware diagnostics:** Runs NVIDIA DCGM diagnostics and targeted stress and power checks to look for faults.
- **Compute and bandwidth benchmarks:** Measures GEMM throughput, GPU memory bandwidth and host-to-GPU PCIe transfer speed.
- **Multi-node collective tests:** Cluster profiles run NCCL all-reduce across nodes over RDMA.
- **Threshold-based reports:** Each measurement is compared with its profile's acceptance threshold and can be saved as a PDF or shared by link for 30 days.
- **Temporary test files:** Test files are copied in for the run and removed, with no packages or services installed.

## Good for

- Acceptance testing newly delivered NVIDIA H100 or B200 nodes before production use
- Checking GPU node health after repairs or driver changes
- Producing a shareable performance report for a GPU cluster handover
- Comparing measured GEMM and bandwidth figures against reviewed thresholds

## Install

### Claude Code (Terminal)

```sh
claude mcp add --transport http gcat https://api.liquidcompute.co/mcp
```

### Claude Desktop (Settings → Connectors → Add custom connector)

```sh
https://api.liquidcompute.co/mcp
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "gcat": {
      "url": "https://api.liquidcompute.co/mcp"
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "gcat": {
      "type": "http",
      "url": "https://api.liquidcompute.co/mcp"
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "gcat": {
      "serverUrl": "https://api.liquidcompute.co/mcp"
    }
  }
}
```

### Grok Bot (Grok Bot desktop app → Connect apps → Add a custom connector)

```sh
https://api.liquidcompute.co/mcp
```

Enter a Name (for example "Liquid Compute GPU test"), choose Type: Remote HTTPS, paste this as the Server URL and choose Add plugin. When we checked, this server asked for OAuth sign-in. If its docs give you an OAuth client ID and secret, enter them in the form's OAuth Client ID and Client Secret fields.

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "gcat": {
      "type": "http",
      "url": "https://api.liquidcompute.co/mcp"
    }
  }
}
```

## FAQ

### Do I need an account?

The website says no account is needed to start a test, follow it and read the report. The README says the MCP server requires a Liquid Compute account sign-in.

### Which GPUs are supported?

The website lists reviewed tests for NVIDIA A10, H100, H200, B200 and B300 hardware. Consumer GeForce RTX GPUs are not currently supported.

### What access does the test need on my nodes?

You add the provided SSH key to authorized_keys for a user with root or passwordless sudo. You can remove that line afterwards.

### How long does a test take?

Duration depends on the configuration. Full diagnostics may take 90 minutes or longer, and the selected configuration shows its maximum duration.

### How do I add Liquid Compute GPU test to Claude Code?

Run this command in your terminal: claude mcp add --transport http gcat https://api.liquidcompute.co/mcp

### What are the best alternatives to Liquid Compute GPU test?

The closest alternatives to Liquid Compute GPU test are XportL, GPU Cloud, CloudGPU MCP, CloudGPU MCP and NVSentinel Lab.

## Where this comes from

- Description: written from the project's README, the project's website and its MCP Registry entry.
- Status: checked automatically by AIPoweredBot on 19 November 2026.
- Install steps: from the official MCP Registry entry.

## Alternatives

- [XportL](https://ai-powered.com/mcp/xportl): XportL is a free MCP server for cloud and infrastructure.
- [GPU Cloud](https://ai-powered.com/mcp/gpu-cloud): GPU Cloud is a paid MCP server for cloud and infrastructure.
- [CloudGPU MCP](https://ai-powered.com/mcp/cloudgpu-mcp): CloudGPU MCP is a paid MCP server for cloud and infrastructure.
- [CloudGPU MCP](https://ai-powered.com/mcp/cloudgpu-mcp-2): CloudGPU MCP is a paid MCP server for cloud and infrastructure.
- [NVSentinel Lab](https://ai-powered.com/mcp/nvsentinel-lab): NVSentinel Lab is an open-source MCP server for cloud and infrastructure.
- [Xcelsior GPU Cloud](https://ai-powered.com/mcp/xcelsior-gpu-cloud): Xcelsior GPU Cloud is a paid MCP server for cloud and infrastructure.
- [EMT Compute](https://ai-powered.com/mcp/emt-compute): EMT Compute is a paid MCP server for cloud and infrastructure.
- [AgentCompute](https://ai-powered.com/mcp/agentcompute): AgentCompute is a free MCP server for design and media.

---
Source: https://ai-powered.com/mcp/liquid-compute-gpu-test
