Rent a GPU From Inside Cline: the ParalonCloud MCP Server in VS Code

The moment you most want a GPU is the moment you are already in the editor: the training script is written, the notebook is open, and the laptop has an integrated graphics chip. Cline speaks MCP, and ParalonCloud ships an MCP server for its GPU marketplace, so the agent that just wrote the script can also rent the card to run it, wait for it to come up, and give you the JupyterLab link. When you are done, it stops the rental and the meter.

This is the same server we published for Claude Code and Cursor; this post is the Cline setup end to end, for the VS Code and JetBrains extensions and for the Cline CLI.

1. A key with the rental scope

Create a key in the Console. A key is inference-only by default; renting spends credits, so it is a separate, opt-in permission. Tick GPU rentals when you create the key, and set the At most field, the most rentals this key may hold at once. One or two is right for a coding agent.

The same prlc_ key also works with the inference API, which is the point: one key, one balance, and an agent that can call a model and rent the hardware for the heavy part of the job in the same conversation.

2. Add the server to Cline

Cline keeps its MCP servers in cline_mcp_settings.json. In the VS Code and JetBrains extensions, open the MCP Servers panel from the Cline toolbar and choose Configure MCP Servers, which opens that file. For the CLI the file lives at ~/.cline/data/settings/cline_mcp_settings.json. Either way, add:

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

npx fetches the package on first use; there is nothing to clone or build. Save, and the server shows up in the MCP panel with its tools listed. If it shows an error instead, it is almost always the key: check that it has the rental scope and that Node.js is on the path Cline uses.

3. Six tools, the whole lifecycle

Tool	What it does	Spends credits
list_gpus	GPUs available right now, with model, VRAM, CUDA compute capability, price per hour and country	no
get_balance	your credit balance	no
create_rental	start a JupyterLab rental on a node, optionally with an auto-stop after N hours	yes
get_rental	status of a rental and, once the tunnel is up, its connection URL	no
list_rentals	what this key has running	no
destroy_rental	stop a rental and its billing	ends it

Renting is asynchronous and the tools say so: create_rental returns at once with status: pending, and the agent polls get_rental until connection_pending is false, usually a few seconds. You watch Cline start the GPU, wait, and paste a working link. It cannot invent one; the URL only exists after the tunnel is up.

4. A session

In Cline, with the server connected:

Find me the cheapest verified GPU with at least 24 GB of VRAM and CUDA compute capability 8.9 or higher, rent it for 2 hours as JupyterLab, and give me the link when it is ready.

Cline calls list_gpus, picks a node, and stops to ask before calling create_rental: that call is billed, and Cline confirms tool calls that change things unless you have auto-approved them. Say yes, and it polls get_rental and returns the JupyterLab URL with its token. Upload the script, run it on the card, and when you are done:

Stop the rental.

destroy_rental, and the meter stops. If you forget, the 2-hour auto-stop stops it for you.

From the CLI, the same thing in one line, in plan mode so nothing is billed before you approve:

cline -p "rent the cheapest verified 24 GB GPU for 2 hours as JupyterLab and give me the link"

5. What keeps an agent from spending your money

An agent with a payment method needs limits, and these are built in rather than left to the prompt:

• Confirmation. create_rental and destroy_rental are the calls that change what you are billed. Cline asks before running them. Do not auto-approve this server.
• Idempotency. A retried create_rental returns the same rental instead of starting a second GPU.
• Auto-stop. Pass a number of hours and the rental ends itself, server side, even if the agent, the editor or the laptop is gone by then.
• Scoped, capped keys. The rental scope is off by default, and the key's At most cap bounds how many GPUs it can hold. A leaked key cannot run up a fleet.
• Locked price. The per-hour rate you saw in list_gpus is fixed for the life of that rental.

6. What it costs

Billing is per minute while the rental runs, at the price the node's owner set, shown per hour in list_gpus. ParalonCloud is a marketplace of independent GPU owners, so the same card is offered at different prices; list_gpus sorts that out for the agent. There is no minimum, no reservation fee and nothing to cancel: a rental you destroy after eleven minutes costs eleven minutes. Credits are prepaid, from $5 by card or crypto, in Add Credits.

Troubleshooting

Symptom	Fix
The server shows an error in the MCP panel	the key lacks the GPU rentals scope, or Node.js is not on Cline's path
create_rental refused with a scope error	same: enable the rental scope on the key in the Console
create_rental refused with a limit error	the key is at its At most cap; stop a rental or raise it in the Console
Insufficient credits	get_balance shows what is left; top up in Add Credits
get_rental keeps returning connection_pending: true	the tunnel is still coming up; the agent should keep polling for a minute before giving up
Cline ran create_rental without asking	you auto-approved the server's tools; turn that off for this one

The server is the same for every MCP client. If you use Claude Code or Cursor, the earlier post has the one-line install for those. And if what you want is a model behind Cline rather than a GPU under it, that is the other setup.