The Model Context Protocol is an open standard that lets AI assistants call external tools. Claude, Cursor, VS Code and many other clients support it.

No. You pay only for the servers the agent runs, at the same per-minute prices as in the dashboard.

Can I use the REST API instead?

Yes, the MCP tools are a thin layer over the same REST API and the same token.

MCP server for AI agents: rent GPUs from Claude, Cursor and VS Code

Address and token

Endpoint: https://vramfarm.com/mcp (Streamable HTTP). Authorization: Bearer gf_live_…, an API token from dashboard settings. The token can only list machines, launch, check and stop servers: it cannot change your password, keys or payments.

Connect your client

Claude Code

claude mcp add --transport http vramfarm https://vramfarm.com/mcp \
  --header "Authorization: Bearer gf_live_YOUR_TOKEN"

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "vramfarm": {
      "url": "https://vramfarm.com/mcp",
      "headers": {
        "Authorization": "Bearer gf_live_YOUR_TOKEN"
      }
    }
  }
}

VS Code (.vscode/mcp.json)

{
  "servers": {
    "vramfarm": {
      "type": "http",
      "url": "https://vramfarm.com/mcp",
      "headers": {
        "Authorization": "Bearer gf_live_YOUR_TOKEN"
      }
    }
  }
}

Claude Desktop and other stdio clients (via mcp-remote)

{
  "mcpServers": {
    "vramfarm": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://vramfarm.com/mcp",
        "--header",
        "Authorization: Bearer gf_live_YOUR_TOKEN"
      ]
    }
  }
}

What the agent can do

`list_gpu_types`	Every GPU model with VRAM, the cheapest price per hour and how many machines are free
`list_offers`	Concrete machines of one model: country, CUDA, disk, network, price
`launch_server`	Rent a machine with a template (PyTorch, ComfyUI, chat, vLLM, text to speech) or your Docker image. Needs max_eur_per_hour, the agent’s budget
`get_server`	Status, cost so far, SSH command, Jupyter and app links
`list_servers`	Your running and recent servers
`stop_server`	Stop billing; optionally save /workspace for the next launch
`get_account`	Balance and how long it lasts

Spending stays under control

Every launch carries max_eur_per_hour: a machine above the agent’s budget is refused before anything is rented.
The balance is prepaid: an agent can never spend more than you topped up.
Billing is per minute after a 15-minute minimum, and idle servers get an email and auto-stop.
Revoke the token in the dashboard at any time; the agent loses access at once.

Try asking your agent

“Find the cheapest GPU with at least 24 GB under €0.50 an hour and start PyTorch on it.”
“Launch an RTX 4090 with ComfyUI and Flux, give me the link, and stop it in an hour.”
“Run my training script from GitHub on an A100, watch it over SSH and stop the server when it finishes.”

MCP server: let your AI agent rent the GPU