orl 0.4: GPU pods, VM sizing, and spend in the OpenRelay CLI

Sep 26, 2026 · 3 min read

orl is the OpenRelay command-line tool. Version 0.4 runs your containers on whole GPUs, checks a VM's size against the fleet before launch, and shows what each VM costs while it runs.

Install and log in

brew install openrelayinc/tap/orl
orl auth login

On Linux, install the .deb or .rpm from the release page. Already installed? Run orl upgrade.

Rent a GPU VM

orl ssh-keys create --name laptop --public-key "$(cat ~/.ssh/id_ed25519.pub)"
orl ssh-keys list
orl deploy trainer --ssh-key <key-id> --connect

orl deploy lists the GPU models with free capacity, asks for a model and count, and checks any --guest-mem-mb or --disk-gb you set against the fleet before it submits. --connect waits for the VM and opens an SSH session.

Run a container on a GPU

orl pods capacity list
orl pods create --image pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime \
  --gpu-model-id <gpu-model-id> --ssh-key-ids <key-id> --volume-gb 100
orl vms metrics get <pod-id> --range 1h

A pod runs your image on whole GPUs, with SSH, an HTTP endpoint, and a /workspace volume that survives stop and start. Use a CUDA image on NVIDIA or a ROCm image such as rocm/pytorch on AMD, and read GPU utilization, memory, power, and temperature with vms metrics get.

Watch spend, stop billing

orl vms burn get <vm-id>
orl usage daily get --days 7
orl vms stop <vm-id>

burn get shows a VM's hourly rate, the cost of its current session, and your runway. usage daily get splits daily spend into compute and inference, and vms stop stops the meter while keeping the disk.

Run batch inference

orl batches create --input-file requests.jsonl \
  --endpoint /v1/chat/completions --priority expedited
orl batches get <batch-id>
orl files content get <output-file-id> -o results.jsonl

expedited gets a slot ahead of waiting standard batches and bills at the online rate. standard keeps the batch discount, and the Batch API is enabled per organization.

Give it to an agent

claude mcp add openrelay -- orl mcp serve

orl mcp serve exposes a core set of OpenRelay API calls as MCP tools. Widen it with --toolset vms, or add --read-only for an agent that can look but never change anything.