Overview

This MCP (Model Context Protocol) server exposes Hyperbolic’s GPU cloud to agents and LLMs so they can discover, rent, SSH into, and run GPU-accelerated workloads. It acts as a local bridge between an LLM-enabled client (for example, Claude for Desktop) and the Hyperbolic API, translating user intents into actions like listing available GPUs, creating a GPU instance, or executing remote shell commands on an allocated node.

For developers building agent workflows, research pipelines, or automated orchestration, the server simplifies common GPU-cloud tasks by packaging them as tools that an LLM or programmatic agent can invoke. It keeps credentials local, supports SSH access to allocated nodes, and surfaces cluster and instance metadata so agents can make informed decisions about renting and using GPU resources.

Features

List available GPUs and cluster details from the Hyperbolic GPU cloud
Rent and terminate GPU instances programmatically via the Hyperbolic API
SSH into rented instances using a provided private key and run remote commands
Tools oriented for LLM/agent use through the Model Context Protocol (MCP)
Local-only execution: credentials remain on the host running the MCP server
Simple integration with LLM clients (example: Claude for Desktop)

Prerequisites

Node.js 16+ and npm or yarn
Hyperbolic account with an API token and sufficient credits
(Optional but recommended) SSH key pair for connecting to GPU instances
Network access to Hyperbolic API endpoints

Installation / Configuration

Clone and install:

git clone https://github.com/HyperbolicLabs/hyperbolic-mcp.git
cd hyperbolic-mcp
npm install
npm run build

Environment variables can be supplied via a .env file or injected by your LLM client. Example .env:

HYPERBOLIC_API_TOKEN=sk_live_....
SSH_PRIVATE_KEY_PATH=/home/you/.ssh/id_rsa
PORT=3000

Or embed in a Claude for Desktop MCP config:

{
  "mcpServers": {
    "hyperbolic-gpu": {
      "command": "node",
      "args": ["/path/to/hyperbolic-mcp/build/index.js"],
      "env": {
        "HYPERBOLIC_API_TOKEN": "sk_live_....",
        "SSH_PRIVATE_KEY_PATH": "/home/you/.ssh/id_rsa"
      }
    }
  }
}

Run the server locally:

npm start
# or
node build/index.js

Note: Private keys should not be passphrase-protected because the server does not handle interactive passphrase prompts.

Available Tools

The MCP server exposes tool-like endpoints for agent use. Summary table:

Tool	Purpose	Key parameters
list-available-gpus	Enumerate GPUs and cluster availability	filter (optional)
get-cluster-details	Fetch metadata about a cluster	cluster_name
rent-gpu-instance	Create/rent a GPU instance	cluster_name, node_name, gpu_count
list-user-instances	List active instances for your account	none
terminate-gpu-instance	Terminate a rented instance	instance_id
ssh-connect	Establish SSH session to an instance	host, username, private_key_path
remote-shell	Execute a command on the connected host	command
ssh-status	Report connection state	none
ssh-disconnect	Close SSH session	none

Each tool accepts structured parameters so LLMs and agents can call them deterministically via MCP.

Use Cases

Rapid experimentation: Ask an agent to “Find a cluster with 4 x A100 GPUs and rent it for 2 hours,” then automatically run a training job via remote-shell commands.
Interactive model debugging: Rent a small GPU node, SSH into it, run diagnostics like nvidia-smi and top, and stream results back to the agent for analysis.
Cost-aware orchestration: Query available GPUs, compare cluster pricing/availability, rent the most cost-effective configuration, and terminate the instance when a job completes.
Multi-agent workflows: Use one agent to provision GPUs and another agent to deploy evaluation workloads, with the MCP server coordinating instance lifecycle and SSH access.

Concrete example (agent prompt -> action):

Prompt: “List all available GPUs.”
Tool: list-available-gpus -> returns clusters and GPU counts.
Prompt: “Rent 2 GPUs from cluster extrasmall-foo for training.”
Tool: rent-gpu-instance -> returns instance_id and host.
Prompt: “SSH into the instance and run nvidia-smi.”
Tools: ssh-connect -> remote-shell -> ssh-disconnect

Security & Billing Notes

The server requires your Hyperbolic API token; treat it like any API secret.
GPU rents incur charges on your Hyperbolic account. Verify funds before provisioning instances.
Keep your SSH private key secure and avoid storing it in world-readable locations.
The server runs locally and does not upload your secrets to external services beyond Hyperbolic’s API.

Troubleshooting

Invalid token or authentication errors: confirm HYPERBOLIC_API_TOKEN and verify account status.
Insufficient credits: top up in the Hyperbolic dashboard under Billing.
SSH failures: check private key path, file permissions (600), and that the instance is reachable on the expected port.
Logs: inspect the server output for API errors and stack traces.

For source code, issues, and contributions, see the GitHub repository: https://github.com/HyperbolicLabs/hyperbolic-mcp.

Hyperbolic MCP Server for GPU Cloud Rentals