Prime Intellect CLI

Prime Intellect CLI

The prime CLI is the command-line interface for managing Prime Intellect compute resources, RL environments, code sandboxes, and tunnels. This skill provides the command reference and decision patterns for all platform operations.

Prerequisites

• Python 3.10+
• Install: uv tool install prime (preferred) or pip install prime
• Auth: prime login (interactive) or prime config set-api-key (headless)
• SSH key: prime config set-ssh-key-path (for pod access)
• Verify: prime config view

Command Reference

Configuration (prime config)

GPU Availability (prime availability)

# List all GPU configurations with pricing
prime availability list

# Filter by GPU type, region, count, socket
prime availability list --gpu-type H100_80GB --regions united_states --gpu-count 2 --socket PCIe

# Find GPUs compatible with existing disks
prime availability list --disks disk-id-1 --disks disk-id-2

# List disk availability
prime availability disks --regions united_states

Filters: --gpu-type, --gpu-count, --regions, --socket (PCIe, SXM2, SXM3, SXM4, SXM5), --disks, --group-similar (default true).

Pod Management (prime pods)

# Create interactively
prime pods create

# Create non-interactively
prime pods create \
  --gpu-type H100_80GB \
  --gpu-count 1 \
  --disk-size 100 \
  --name my-pod

# Create with persistent disks attached
prime pods create --id 346663 --disks disk-id-1 --disks disk-id-2

# Create with custom template
prime pods create --image custom_template --custom-template-id "template_id"

# List pods
prime pods list

# SSH into a pod
prime pods ssh <pod-id>

# Delete a pod
prime pods delete <pod-id>

Pod creation options: --id, --cloud-id, --gpu-type, --gpu-count, --name, --disk-size, --vcpus, --memory, --image, --team-id, --env KEY=value, --disks, --share-with-team, --add-members.

Disk Management (prime disks)

# Check availability
prime availability disks

# Create a persistent disk
prime disks create --id c008ad --size 500 --name ml-training-data

# List disks (with pagination)
prime disks list --limit 50 --offset 0 --output json

# Delete a disk
prime disks delete <disk-id>

Disks persist independently from pods and are billed continuously until terminated. Use --yes to skip confirmation in automation.

Sandboxes (prime sandbox)

# Create a sandbox
prime sandbox create python:3.11-slim \
  --name analytics-lab \
  --cpu-cores 2 --memory-gb 4 --disk-size-gb 20 \
  --timeout-minutes 240 --idle-timeout-minutes 15 \
  --env PROFILE=production \
  --secret DB_PASSWORD=<value> \
  --label experiment --label ml-pipeline

# Run a command inside
prime sandbox run sbx_123 --working-dir /workspace "python -c 'print(42)'"

# Upload / download (200MB per-file limit)
prime sandbox upload sbx_123 notebooks/analysis.ipynb /workspace/
prime sandbox download sbx_123 /workspace/report.csv reports/latest.csv

# Expose ports (range 22-9000; 8080, 2222, 8081 excluded)
prime sandbox expose <sandbox-id> 8000 --name web-server
prime sandbox expose <sandbox-id> 9000 --name tcp-server --protocol TCP
prime sandbox list-ports <sandbox-id>
prime sandbox unexpose <sandbox-id> <exposure-id> --yes

# SSH into a sandbox
prime sandbox ssh <sandbox-id> --shell zsh

# Inspect, logs, cleanup
prime sandbox list --status RUNNING --output table
prime sandbox get sbx_123 --output json
prime sandbox logs sbx_123 > logs.txt
prime sandbox delete --label experiment --yes

Idle timeout constraints:

• Disabled by default; opt in with --idle-timeout-minutes.
• Must satisfy 1 <= idle <= timeout and idle <= 1440.
• Not supported for VM-backed sandboxes (--vm).
• SSH sessions do not count as activity.
• File transfer auth errors: prime sandbox reset-cache then retry.

Hosted RL Training (prime lab, prime train, prime eval)

# Set up workspace
prime lab setup

# Install an environment
prime env install primeintellect/alphabet-sort

# Run baseline evaluation
prime eval run primeintellect/alphabet-sort \
  -m Qwen/Qwen3-4B-Instruct-2507 -n 20 -r 1
prime eval tui

# Launch training
prime train run configs/rl/alphabet-sort.toml

# Monitor
prime train logs <run-id> -f
prime train models    # list available models

Training config (TOML):

model = "Qwen/Qwen3-4B-Instruct-2507"
max_steps = 50
batch_size = 128
rollouts_per_example = 8

[sampling]
max_tokens = 512

[[env]]
id = "primeintellect/alphabet-sort"

# Optional W&B integration
[wandb]
project = "my-experiment"
name = "alphabet-sort-30b"

# Optional periodic eval
[eval]
interval = 50

Run size guidelines:

RL Environments Hub (prime env)

prime env list
prime env info owner/environment-name
prime env install owner/environment-name
prime env init my-new-environment

Prime Tunnel (prime tunnel)

# Expose a local service
prime tunnel start --port 8000

# With basic auth (password auto-generated, shown once)
prime tunnel start --port 8000 --auth alice

Returns a public HTTPS URL like https://t-0-abc123def456.tunnel.pinfra.io. Hosted evaluations can use tunnels with --allow-tunnel-access flag.

Common Mistakes

• Forgetting to clean up disks: disks are billed continuously until explicitly deleted, even after pods are terminated. Always prime disks delete when done.
• Using --env for secrets: env vars are plaintext and visible on inspect. Use --secret for credentials.
• Sandbox idle timeout with SSH: SSH sessions do not count as activity. A sandbox can be reaped while you have an active SSH session.
• Port restrictions: ports 8080, 2222, and 8081 cannot be exposed. Valid range is 22-9000.
• 200MB file transfer limit: individual file uploads/downloads are capped at 200MB. Use multiple calls or compress first.
• Not setting team context: if billing to a team, set prime config set-team-id before provisioning, or pass --team-id per command.
• Custom template compatibility: ensure templates are compatible with the selected GPU configuration before creating pods.