MCP
Connect once. Then ask in plain words.
MCP lets Claude Code, Codex or Cursor use outside tools. Add Runcrate with one command and your agent launches GPUs, runs your job on them and stops the bill, from $0.51/hr.
You sign in through your browser. Your agent never holds a key.
Connect
One command, then sign in. No key to copy first.
- 01
Add Runcrate.
One click in Cursor or VS Code, one command in Claude Code or Codex, or the URL in Claude or ChatGPT.
- 02
Sign in when it asks.
Your browser opens the Runcrate sign-in. Approve, and your agent works in your workspace.
- 03
Ask.
In plain words, the way you would ask a person: “Launch an H100 with the Finetune crate and run nvidia-smi.”
claude mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcpTools
47 tools your agent can use. It picks the right one for each request.
Instances
See prices, launch a GPU on a crate, check on it, set when it stops on its own, and stop it.
Jobs
Run a command on its own GPU machine with your code, or a sweep of them, waiting for a free GPU from a list and retrying on a new machine. Read the output, the files it kept, its timeline and its GPU use; billing stops when the command exits.
Schedules
Start an instance or a job later, once or on a repeat, and pause, run or delete the schedule.
Run and files
Run a command on a running instance, and copy small files to and from it, such as a script or a results file.
Storage
Volumes that keep your files when an instance stops.
Billing
Your balance, what you spent on what, and how much of an environment’s monthly budget is left.
Crates and SSH keys
The ready images an instance can boot, and the keys for your own ssh.
Environments and secrets
The environment your agent works in, picked when you connect it, and the names of the secrets its machines get. Environments and secret values are set in the dashboard, never by the agent.
Firewall
Open and close inbound ports for every machine in an environment.
HTTPS links
Serve a port of an instance at its own HTTPS link, for your workspace, for anyone, or for listed addresses.
Examples
Ask the way you would ask a person. Real prompts from the docs, and the tools they call.
“Deploy an A100, upload my dataset, and start training.”
“Compare A100, H100, and RTX 4090 pricing. I need at least 40GB VRAM.”
“Set up a vLLM server with Llama 3.1 70B on an A100.”
“My training script is crashing with CUDA OOM. Can you check what's happening?”
“Find anything that's running but not being used and shut it down.”
Sign-in and safety
You sign in. Your agent never holds a key.
A browser sign-in, never a pasted key
Your client registers itself and opens the Runcrate sign-in page. Your password never passes through the agent, and no key goes in a config file.
Every tool says what it does
Each tool is marked read-only, creating, or deleting, so your client can ask you before your agent stops an instance or deletes a volume.
Your workspaces, your environments
Your agent acts only on workspaces you belong to, and every tool can be pointed at one environment, such as staging.
100 requests a minute
Far more than a person asking needs, and a stop on a runaway loop.
Pricing
The prices in your dashboard. Your agent pays the same rate you do.
Billed by the second on most GPUs while an instance runs; a few bill by the minute. Delete it and the bill stops.
FAQ