MCP server (Claude, Cursor)

The ParalonCloud Rentals MCP server exposes the Rental API to any MCP client — Claude Code, Claude Desktop, Cursor — so you can rent a GPU, poll it for its connection URL, and stop it, all in natural language.

Ask "rent me a 24GB GPU for 2 hours and give me the Jupyter link", and the agent browses GPUs, starts the rental, waits for it to come up, and hands you the URL.

• npm: @paraloncloud/mcp-rentals
• Source: github.com/ParalonCloud/mcp-rentals

Tools

Tool	What it does	Cost
list_gpus	List rentable GPUs with price, VRAM, compute capability, country	free
get_balance	Your credit balance	free
create_rental	Start a rental (async → poll)	spends credits
get_rental	Status + connection URL once running	free
list_rentals	Your active rentals	free
destroy_rental	Stop a rental and stop billing	—

Get a key first

The MCP needs a prlc_ API key with the rental scope:

1. Create a key in the Console.
2. Turn on the GPU Rentals scope for it (rentals are opt-in — see Authentication & scopes).
3. Use a dedicated key for the agent, not your production key.

Install

The client passes your key via the PARALON_API_KEY environment variable — you never edit the server.

Claude Code

claude mcp add paraloncloud-rentals \
  --env PARALON_API_KEY=prlc_your_key_here \
  -- npx -y @paraloncloud/mcp-rentals

Verify with claude mcp list (should show ✓ connected), then /mcp to see the tools.

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "paraloncloud-rentals": {
      "command": "npx",
      "args": ["-y", "@paraloncloud/mcp-rentals"],
      "env": {
        "PARALON_API_KEY": "prlc_your_key_here"
      }
    }
  }
}

Restart Claude Desktop.

Cursor

Add the same block to .cursor/mcp.json (project) or the global Cursor MCP config.

Optional env: PARALON_BASE_URL (defaults to https://paraloncloud.com/api/v1).

Safety

Renting spends real credits, so the server is built to be safe to hand to an agent:

• Confirmation. create_rental and destroy_rental change what you're billed — MCP clients confirm tool calls by default. Keep that on.
• Idempotency. create_rental auto-generates an idempotency key, so a retried call never starts a second GPU.
• Auto-stop. Pass hours so a rental stops itself even if the agent forgets.
• Concurrency cap. The key's max_active_rentals limit (set in the Console) bounds how many rentals it can hold at once.

How it feels

Renting is asynchronous, and the tools tell the agent so: create_rental returns immediately, then the agent polls get_rental until connection_pending is false (the secure tunnel comes up a few seconds after the rental starts). You'll see it start the GPU, wait, and then give you a working link.

One key does both: the same key powers the Inference API. Build an agent that calls a model and rents the GPU to run the heavy job.