Check CUDA before committing.
GPU environment inspection on an RTX 3090. A machine-readable environment report.
191stask elapsed
$0.0048final GPU fee
Agent Compute Lab / No account required
Compute for agents.Human control.
Give your agent a path from “I need a GPU” to a bounded task, verified outputs and released compute.
MCP · REST · Agent Skills · explicit paid-work approval
$ GET /v1/public/agent-self-checkNo login. No API key. No allocation. Check live capabilities, billing constraints, supported task types and known issues.
This check spends $0 and operates no customer GPU. Open raw live check ↗
01 / DISCOVERCheck task fit
Read capabilities, inventory and limitations.
02 / APPROVESet the boundaries
Owner connects and approves the actual quote.
03 / RUNExecute bounded work
Use time and spend caps. Reconcile uncertain starts.
04 / VERIFYExport and release
Check outputs and confirm compute is released.
Live discovery / no account / no allocation
Ask for a real task estimate.
Browse minute-billed candidates. Quotes are estimates, not reservations or permission to spend. Whole-hour offers require a separate binding quote.
No key or payment required. This action creates no GPU.
Costs cover estimated GPU runtime; storage and transfer are separate. Account connection and explicit acceptance of a fresh binding quote are required before paid work. Read task and billing requirements ↗
Interactive / local-only / no rental
What does early stopping cost?
A billing illustration, not a quote. Compare rounding before your agent allocates.
$0.08minute-billed source
$0.40whole-hour source
GPU rates and supply are not fetched here. Storage and other charges are excluded. Actual start conditions and minimums come from the binding quote.
Evidence, not promises
Three small tasks. Real measurements.
Historical internal smoke tests. Fees and time below describe the successful attempt, not a price quote or production success rate.
Inspect runnable recipe ↗
Generate a tiny Wan video.
Wan2.1 1.3B in ComfyUI; 17 H.264 frames at 256 × 256. Development: 1 of 12 attempts succeeded; all attempts cost $0.8775.
1,064stask elapsed
$0.1105successful GPU fee
Train a small language adapter.
SmolLM2-135M causal-LM LoRA, 8 steps. Development: 1 of 4 attempts succeeded; all attempts cost $0.0339. Not image LoRA.
190stask elapsed
$0.0097successful GPU fee
These tests verify narrowly defined recipes, not arbitrary workflows or an SLA. Across the recorded acceptance session, all 19 allocated instances were confirmed released; one stock failure never allocated. This is internal evidence, not a customer reliability statistic.
A closer look / 20-second read-only UI tour
Explore before connecting.
Recorded UI tour. No paid GPU task is started in this video. Read transcript · Technical guide: the hourly price is not your task budget
Bring your own agent
Install the playbook.
Keep spending explicit.
Use the Agent Skills CLI, then configure CloudGPU MCP using your client's documentation. The owner signs in and approves connection limits; paid work still requires approval.
MCP setup and authorization ↗Agent Skills
npx skills add cloudgpu-app/cloudgpu-agent-skillsChoose your skills and client when prompted. Installation starts no GPU.
Remote MCP
https://cloudgpu.app/mcpStreamable HTTP + OAuth for compatible clients. Local stdio: npx -y cloudgpu-mcp. Never paste API keys into chat.
Know before you run
Billing follows the source.
Minutes or whole hours — check the quote.
Minute-billed sources round partial minutes up. Whole-hour sources have a one-hour minimum and charge every started hour in full, even when stopped early. A GPU name alone does not determine billing.
Results need a storage plan.
Instance disks are erased on termination. Configure and verify Job outputs or Workspace storage. Capacity budgets, retention and transfer limits are separate from rental charges.
Machine-readable contract ↗Live limits / no hidden green lights
Readiness is not an SLA.
Open the live self-check to read current known issues and evidence availability. The check reports the control plane; it does not test model generation or guarantee upstream availability.
Read the raw live endpoint. We do not substitute cached results as a live pass.