Liquid Compute GPU test

Liquid Compute GPU test (GCAT)

Run GPU acceptance tests on NVIDIA GPU nodes you administer, from your AI agent.

Liquid Compute connects to your nodes over SSH from its own servers. It runs NVIDIA DCGM diagnostics, a CUTLASS GEMM benchmark, HBM and PCIe bandwidth tests and NCCL all-reduce for a GPU profile, then reports every measurement against that profile's thresholds. Profiles cover A10 nodes, H100 and H200 8-GPU nodes and 2- or 4-node clusters over RDMA, and B200 and B300 8-GPU nodes. Your agent can list the current ones.

The test is free and needs a Liquid Compute account. New to Liquid Compute? Create your account and finish setup on the website first. Your agent asks you to sign in the first time it connects.

This repository holds no code. The hosted MCP server at https://api.liquidcompute.co/mcp does the work; this repository has the listing files for agent marketplaces and an agent skill that teaches agents the test flow.

Requirements

Every node needs:

• Ubuntu 22.04 or 24.04 with NVIDIA driver 580 or later (CUDA 13.0 or later);
• a public IP address that accepts SSH from the source addresses Liquid Compute gives you when you create a test;
• a user with root or passwordless sudo;
• idle GPUs for the whole test. A test puts every GPU under full load for the profile's duration, which can be 90 minutes or longer.

Connect your agent

The server uses Streamable HTTP and OAuth 2.1 sign-in with the gpu_tests scope.

Claude Code

Install the plugin, which adds the server and the skill:

claude plugin marketplace add Liquid-Compute/gcatclaude plugin install liquidcompute-gcat@liquidcompute

Or add only the server:

claude mcp add --transport http --scope user liquidcompute-gpu-test https://api.liquidcompute.co/mcp

Then run /mcp in Claude Code to sign in.

Claude apps

In Claude, open Customize > Connectors, select Add custom connector and enter https://api.liquidcompute.co/mcp. On Team and Enterprise plans, an Owner adds it under Organization settings > Connectors.

Codex

codex mcp add liquidcompute-gpu-test --url https://api.liquidcompute.co/mcpcodex mcp login liquidcompute-gpu-test

ChatGPT

Open chatgpt.com/plugins, select +, then Add custom MCP server. Enter https://api.liquidcompute.co/mcp and choose OAuth. Your workspace settings decide whether custom MCP servers are allowed.

Cursor

Open this install link in your browser:

cursor://anysphere.cursor-deeplink/mcp/install?name=liquidcompute-gpu-test&config=eyJ1cmwiOiJodHRwczovL2FwaS5saXF1aWRjb21wdXRlLmNvL21jcCJ9

Or add the server to ~/.cursor/mcp.json:

{  "mcpServers": {    "liquidcompute-gpu-test": { "url": "https://api.liquidcompute.co/mcp" }  }}

VS Code

Run MCP: Add Server, choose HTTP and enter https://api.liquidcompute.co/mcp. Or add it to .mcp.json at your workspace root:

{  "mcpServers": {    "liquidcompute-gpu-test": { "type": "http", "url": "https://api.liquidcompute.co/mcp" }  }}

Gemini CLI

Install the extension, which adds the server and the skill:

gemini extensions install https://github.com/Liquid-Compute/gcat

Or add only the server, then sign in with /mcp auth liquidcompute-gpu-test:

gemini mcp add --transport http --scope user liquidcompute-gpu-test https://api.liquidcompute.co/mcp

How a test runs

1. Your agent picks the profile that matches your nodes and creates a test. You get one authorized_keys line and the addresses Liquid Compute connects from.
2. You add the line for a root or passwordless-sudo user on every node, or let your agent add it over SSH access you gave it.
3. You confirm that you are authorized to give Liquid Compute SSH access. Your agent adds the nodes, and Liquid Compute runs a brief connection check that does not load the GPUs.
4. You confirm that the nodes are reserved, their GPUs will stay idle, and you are authorized to run the test. The test starts.
5. Your agent follows the test and reads you the report. It can also create a link to the report that anyone with the link can open for 30 days.
6. You remove the authorized_keys line.

The skill has the full flow and the tools it uses.

A report measures your nodes during the test window. It is not a certification or a warranty. If your GPUs pass and you own them, you can list them for rent on Liquid Compute.

Safety

• Two confirmations only you can give. The server refuses to add nodes or start a test until your agent passes your own confirmation (steps 3 and 4). The skill tells agents to show you the exact wording and never to confirm for you. Don't set your client to auto-approve these two tools.
• Annotated tools. Every tool says whether it only reads, and starting and stopping a test are marked destructive, so your client can ask before running them.
• Nothing installed. Test files are copied to the nodes for the test and removed afterwards. Liquid Compute installs no packages, services or accounts and changes no settings.
• Access you control. Liquid Compute can reach a node over SSH until you remove its authorized_keys line.

Data

• Sign-in uses your Liquid Compute account. The token your agent receives is issued for this server only.
• Liquid Compute stores the tests you create, the node addresses and user names you add, the hardware details and raw output the test collects, connection-check results and test reports, under your company's account, until you ask for them to be deleted.
• Reports stay available until the expiry shown when you create a test. Share links last 30 days.

Liquid Compute's Terms of Service and GPU test privacy notice apply. The notice lists what a test collects, how Liquid Compute uses it, how long it is kept and how to ask for it to be deleted.

Support

Open an issue at github.com/Liquid-Compute/gcat/issues for help with the server, the skill or a test. Issues are public, so don't post node addresses, keys or report links in them. Send privacy requests to [email protected].

License

The listing files and skill text in this repository are under the MIT License. Use of the Liquid Compute service is governed by its Terms of Service.