▸ ephemeral gpu sessions for coding agents

Your agent writes CUDA.
Now it can run it.

One command gives Claude Code a real GPU. No AWS account, no SSH keys, no setup.

boot time < 60s
billing per-second
setup one command
~/cuda-kernels · claude ● live

▸ why gpus r us

Everything your agent needs.
Nothing you don't.

01

No cloud account needed

Runs entirely in our infrastructure. No AWS, no IAM roles, no VPCs. Your agent gets a GPU shell — you don't touch a thing.

02

Ready in under a minute

nvcc, the CUDA toolkit, and accelerated libraries — pre-built and ready to compile. No driver installs, no environment setup, no wasted time.

03

Only pay for what runs

Per-second billing. Your agent can iterate on the same session all day — you're only charged for actual GPU time.

04

Results land in your repo

When the work is done, your agent opens a pull request with the results. Your code, your branch, zero manual steps.

▸ how it works

From prompt to GPU.

You describe the task. Your agent handles the rest — provisioning, iterating, and cleaning up when done.

01

Install in 30 seconds

One command installs and configures everything. Your agent picks up the GPU tools automatically — no config files, no manual steps.

curl -fsSL https://gpusr.us/install.sh | bash

02

Your agent spins up a session

Describe the task in plain language. Your agent picks the right environment and provisions a GPU — ready in under a minute.

provision_gpu("cuda-developer")

03

Iterate freely, then terminate

Your agent compiles, benchmarks, reads the output, and retries — all in one session. When done it terminates and you're billed only for what ran.

exec(id, "nvcc -O3 kernel.cu && ./bench")

▸ runtime environments

Ready when Claude is.

Pre-built Docker images on NVIDIA-optimized instances. No setup, no driver installs — containers start and run.

cuda-developer

CUDA Developer

Full CUDA toolkit with nvcc, CMake, GDB, and build essentials. For writing, compiling, and profiling custom CUDA kernels on real hardware.

nvidia/cuda:12.4.1-devel-ubuntu22.04
nvcc cuda-toolkit gcc / g++ cmake gdb
ml-runtime

ML Runtime

CUDA-accelerated libraries ready to go — PyTorch, RAPIDS, and Jupyter from the NVIDIA NGC stack. Verify your GPU code paths actually hit the GPU.

nvcr.io/nvidia/pytorch:24.12-py3
PyTorch 2.x RAPIDS Jupyter CUDA 12.4 cuDNN

▸ pricing

Pay for what you use.
Nothing more.

Buy credits upfront, spend them as your agents run. No subscriptions, no minimums. Credits never expire.

starter

$10

10 credits

  • ~24 hrs of g4dn.xlarge
  • All environments
  • Session logs & dashboard

team

$100

115 credits +15 free
  • ~276 hrs of g4dn.xlarge
  • All environments
  • Session logs & dashboard
  • Priority provisioning
  • Email & Slack support

▸ faq

Common questions.

Do I need an AWS account? +

No. GPU instances run entirely in our infrastructure. You never provide cloud credentials, manage IAM roles, or configure networking. We handle all of that — you just use the MCP tools.

How do I install the MCP server? +

Run claude mcp add @gpu-runners/mcp in your terminal. After authenticating with your GPUs R Us account, Claude Code discovers all GPU tools automatically the next time you start a session.

What GPU types are available? +

We currently support the g4dn family on AWS: g4dn.xlarge (1× T4, 16 GB VRAM), g4dn.2xlarge (1× T4, 32 GB), and g4dn.12xlarge (4× T4, 192 GB). More instance types — including A10G and H100 — are on the roadmap.

What happens if a spot instance gets interrupted? +

Transparent spot interruption handling is coming soon. When it ships, your workspace will be automatically checkpointed to a git branch and restored on a new instance — Claude sees a brief delay but never loses context. Today, sessions run on on-demand instances.

Can Claude push code to GitHub? +

Yes. Connect your GitHub account in settings and Claude can use the create_pr MCP tool to open pull requests with results, trained models, or any artifacts from the session.

How does billing work exactly? +

Credits are deducted from your balance as GPU time accrues, billed per second. Your dashboard shows your real-time balance and any charges accruing on active sessions. When a session terminates, the final charge is settled. You can always see your current spend in the top bar.

Is this only for Claude Code? +

The MCP server is purpose-built for Claude Code, but the underlying REST API is open. If you want to integrate GPU provisioning into your own agent or workflow, get in touch.

▸ get started

Ready to give your agent a GPU?

One command sets everything up. Then just describe your task.

terminal
# Install and configure in one step
$ curl -fsSL https://gpusr.us/install.sh | bash

# Then just describe your task to Claude
$ claude
> Write a fused softmax kernel and benchmark it against torch on a T4

yours for as long as the job runs