Skip to main content
Your agent can find GPU and model prices, connect to Runcrate, launch GPUs, run jobs and keep its balance topped up. You sign in once in your browser; there is no key to paste.

​
Connect the MCP server

One command in a terminal. The server is https://mcp.runcrate.ai/mcp: your agent signs in through your browser on first use (OAuth), so there is no key to copy, and it then acts in the one environment you pick, under that environment’s monthly budget.
  • Claude Code: then run /mcp and pick runcrate to sign in.
  • Codex opens your browser as it adds the server. Closed it? Run codex mcp login runcrate.
  • Gemini CLI signs in on first use; /mcp auth runcrate inside it signs in again.
  • Cursor: Add to Cursor in one click (or open cursor://anysphere.cursor-deeplink/mcp/install?name=runcrate&config=eyJ1cmwiOiJodHRwczovL21jcC5ydW5jcmF0ZS5haS9tY3AifQ==), or add {"mcpServers":{"runcrate":{"url":"https://mcp.runcrate.ai/mcp"}}} to ~/.cursor/mcp.json.
Any other agent that can run a command: paste this line into it, and it adds the server to itself.
Every client on this machine at once, with the Runcrate CLI: runcrate mcp install. Claude, ChatGPT and every other client: MCP setup.

​
Install the Runcrate skill

The skill teaches your agent to use Runcrate well: check prices and stock first, run one-off work as a job that stops its own billing, give every machine a max runtime, stop what it started, and top up when the balance runs low. It follows the open Agent Skills format and installs into Claude Code, Codex, Cursor, Gemini CLI and many more:
Add -g to install it for every project. The Claude Code plugin and the Gemini CLI extension carry the same skill with the MCP server:

​
Install the CLI

For an agent that works in a terminal without MCP:
Then runcrate login signs in through your browser. See the CLI reference.

​
Prices and stock an agent can read

No key needed. Every figure is read live from the catalog the pricing pages render.
  • runcrate.ai/pricing.md: GPU prices per GPU-hour, on demand, for every GPU in stock, and every model’s price per 1M tokens (or per image, second or minute), with the time it was read.
  • runcrate.ai/gpus.json: every GPU as JSON with its memory, the price per GPU-hour it starts at, offers listed, regions and prepaid terms. A GPU sold out in the last 30 days stays listed at available_options: 0.
  • runcrate.ai/api/models/catalog: every model as JSON with its billed prices.
  • runcrate.ai/llms.txt: the map of the site for agents.

​
Paying

Runcrate runs on prepaid credits, with no quota requests: the listed rate is the rate, for any GPU in stock. A launch the balance cannot cover is refused with insufficient_credits (a model call with insufficient_balance), both HTTP 402. Your agent reads the balance and runway with billing_balance (CLI: runcrate billing balance) and adds credit from the saved card with top_up (CLI: runcrate billing topup <amount>), once a workspace owner or manager turns on Top-ups for its sign-in. See Top up by API. No credit card? A person can top up without one in the dashboard, under Billing, then Add credits.