GPU credits
Top up NVIDIA compute credits for training and inference on accelerated hardware. The agent funds the GPU time a run demands without a human touching a billing console.
Compute on demand
GPU and cloud credits, bought by the agent
Give an AI agent one API to top up cloud and GPU compute the moment a job needs it. It buys AWS, Google Cloud and NVIDIA credits programmatically, inside per-agent spend caps and human approval, with no shared corporate card in the loop.
AWS · GCP · NVIDIAcredits in one API
190+countries served
Capsenforced on every buy
What agents can fund
The compute an agent needs, on its own budget
Provisioning, not just paying. When a workload runs low, the agent tops up the exact credits it needs and keeps working, all through the same call it uses for every other purchase.
Cloud credits
Add balance to AWS and Google Cloud so pipelines, containers and storage keep running. One API covers the hyperscalers an agent already builds on.
Training & inference budgets
Fund a fine-tune, a batch job or a burst of serving from a budget you set per agent. Spend maps to a task, so every credit purchase is tied to the work that needed it.
Under policy
Buy compute under policy
One call buys credits. Money is integer minor units, so amount_minor: 25000 is $250.00, never a float. A client_ref makes the buy idempotent, and anything over your threshold parks for human approval instead of charging.
# Agent tops up GPU compute for a training run res = client.buy_giftcard( brand="nvidia-compute", amount_minor=25000, # $250.00, integer cents client_ref="train-run-8842", # idempotent: never double-charges ) if res.status == "APPROVAL_REQUIRED": # over the cap: parked, a human is pinged in Slack print(res.hint) elif res.status == "COMPLETED": code = client.get_code(res.task_id) # single-read credit code redeem(code)
- Spend caps. Per-agent daily and monthly limits plus a brand allowlist bound how much compute an agent can buy.
- Human approval. Buys above your configured threshold park and ping a human in Slack before any money moves.
- Fail closed. If the policy engine is unreachable the purchase is denied, never quietly allowed.
- Idempotent. Reuse a
client_refand you get the existing task back. One credit purchase per task, never two. - Full audit. Every buy reconstructs from an OpenTelemetry trace and a double-entry ledger that always sums to zero.
Built for autonomous workloads
Compute spend that stays in your control
Autonomous jobs move fast and unpredictably. The refill layer lets them scale their own compute while every dollar stays inside limits you set and a trail you can read.
Burst capacity
When a run needs more GPUs now, the agent tops up in seconds instead of waiting on a procurement ticket. No idle over-provisioning to cover the spike.
No shared card
Agents never see a corporate card number. Each one authenticates with its own scoped API key and draws on a pre-funded balance, not your raw payment credentials.
Budget you control
Set daily and monthly caps, an allowlist and an approval threshold per agent. Change them anytime, and hit the kill switch to freeze all compute spend at once.
Full audit trail
Every compute buy is one task with a trace and a balanced ledger entry. Reconstruct who spent what, when and why, down to the credit that funded a single run.
Let your agents fund their own compute
Wire up one API and give agents cloud and GPU credits on demand, inside caps you set. Start free in sandbox, no real funds needed.