Run Your GPU Nodes From Claude Code: the Paralon MCP Server and Skills

The provider API lets you run your GPU machines from code. Sometimes you do not want to write code at all; you want to ask. "How are my nodes?", "Is anything running hot?", "Take the second rig off the market until tomorrow." This guide sets that up in Claude Code, and shows a real session.

Two parts: tools and know-how

The Paralon plugin for Claude Code installs two things.

The MCP server (@paraloncloud/mcp) gives Claude Code one tool per API call: list_my_nodes, get_my_node, add_node, get_node_install_command, set_node_available, set_node_price, update_node, reclaim_spot_node and get_earnings for your nodes, plus the six rental tools (list_gpus, create_rental, and the rest).

Three skills teach the agent how to use those tools well. A skill is a short instruction file Claude Code reads when your question matches it:

Install

claude plugin marketplace add ParalonCloud/paralon-skills
claude plugin install paraloncloud@paraloncloud

The first command adds our plugin catalogue, which is a public GitHub repository (ParalonCloud/paralon-skills); the second installs the plugin from it. The server reads your key from the environment, so set it where Claude Code starts:

echo 'export PARALON_API_KEY=prlc_your_key' >> ~/.bashrc
source ~/.bashrc
claude

For your nodes, the key needs the My nodes permission (Console → API keys); for renting, GPU rentals. Give the agent a key of its own with only what it needs. Inside Claude Code, /mcp shows the paraloncloud server as connected and /skills lists the three skills.

A real session

We asked, in plain words, how our nodes were doing. Claude Code picked the gpu-provider skill on its own, called list_my_nodes, and answered:

Then, the part the skill is for, it said what needed attention without being asked:

• The 4090 is fine: 68 °C at 90% load is a good temperature, and the running rental keeps the $0.52 it started with while new rentals would pay the $0.54 default.
• The Mac machines had not connected since May but were still marked available on the market, so they could never be rented while offline: take them off the market, or delete them from the dashboard if they are not coming back.
• A node added recently had never been started: delete it if it was a test (deleting is on the site only, not in the API).

And it stopped there and asked whether to take the offline machines off the market. We said no, and nothing changed. That is the rule the skill gives it: confirm before any change, and never change a rented node under its renter without a clear instruction.

Things to ask

• "How are my nodes? Is any GPU above 80 °C?"
• "Take rig-2 off the market until I say so." (set_node_available; a running rental continues)
• "Is my 4090 priced right?" (it compares with what renters pay for the same GPU via list_gpus, checks the node's allowed range, proposes a price, and waits)
• "Add a new machine called rig-5, Linux, GPU." (add_node; it hands you the install command and warns that the command holds the node's secret token)
• "How much did my nodes earn this week, and which one earned most?" (get_earnings)

What the agent cannot do

The key sets the hard limits, whatever the agent decides. A key with only My nodes cannot rent GPUs, call models or spend credits, and no key can request a payout, change where payouts go or delete a node: those stay on the site. On a spot node, reclaim_spot_node stops a renter's spot rental two minutes later, so the skill always asks first and says so.

Other agents

The MCP server works in any MCP client, with the same key: Claude Desktop, Cursor, Cline. Add a server with the command npx -y @paraloncloud/mcp and the PARALON_API_KEY environment variable; the MCP page has the config for each. The skills are plain instruction files in the same repository, so they can be read and adapted for any agent that takes instructions.

What it costs

The plugin is free and open source (MIT). In Claude Code the three skill descriptions add about 345 tokens to each session; a skill adds 1,000 to 1,600 tokens when it is used. The API calls behind the tools are free. Renting a GPU through create_rental is billed by the minute, and the rent-gpu skill shows the price and waits for your yes before it starts one.