Docs

Command-line interface

Create an API key under API keys and webhooks, then:

rgpu login rgk_live_…
rgpu search --gpu RTX_4090,RTX_5090 --min-reliability 0.95 --sort perf-per-dollar
rgpu launch --offer off_… -t jupyter --disk 50 --wait
rgpu ssh inst_…              # SSH through the relay
rgpu exec inst_… nvidia-smi  # no SSH key needed
rgpu logs inst_… -f
rgpu stop inst_…             # disk kept, GPU reserved for 1 hour
rgpu billing balance

REST API

Base URL https://api.rent-gpu.app/v1. Authenticate with Authorization: Bearer rgk_…. Money is integer micro-dollars (1,000,000 = $1.00). Errors are RFC 9457 problem documents.

POST /v1/offers/search
{"filters": {"gpu_name": {"in": ["RTX 4090"]}, "reliability": {"gte": 0.95}, "bw_down_mbs": {"gte": 500}},
 "order": [["perf_per_dollar", "desc"]], "limit": 20, "pricing": {"disk_gb": 50}}

POST /v1/instances            (Idempotency-Key header supported)
{"offer_id": "off_…", "template_id": "pytorch", "disk_gb": 50, "max_price_micro_hr": 500000}

GET  /v1/instances/{id}       → connect.ssh.command, connect.ports[].url
GET  /v1/instances/{id}/logs?follow=1
GET  /v1/instances/{id}/exec  (WebSocket)
GET  /v1/market/gpu-prices    (public)

Operators: eq ne gt gte lt lte in nin. Unknown fields return 400 with the list of allowed fields.

Verification and reliability

Reliability = uptime² × (1 − 0.05 × failed starts) × (1 − 0.10 × unscheduled downtimes) × (1 − 0.05 × failed proofs) × (1 − 0.10 × incidents), over 30 days. New machines are capped at 0.90, rising to 1.00 over their first 30 days.

If a machine goes down during your rental, we credit you automatically: 5× the hourly price for the downtime (10× on Datacenter hosts), at least $1.