GPU cloud,
from your terminal.
One static binary for macOS, Linux and Windows. Launch a GPU, SSH in, install Jupyter and tear it down without opening a browser. The same REST API runs your CI jobs and your agents.
$ packet launch --gpu b200 --name trainer --wait Launched 1 x NVIDIA B200 Dynamic Instance ID: cmr8x2k9a Price: $4.25/hr Waiting for the instance to take SSH... Ready: ssh -p 30122 ubuntu@203.0.113.24 $ packet setup jupyter-torch cmr8x2k9a $ packet ssh cmr8x2k9a --command "nvidia-smi -L" GPU 0: NVIDIA B200 $ packet terminate cmr8x2k9a --force
Download one file.
That's the install.
No runtime, no package manager. Put the binary on your PATH, paste an API key from the console, and you're in.
Then log in
Create an API key in the console under API keys, and paste it once. In CI, set PACKET_API_KEY instead.
With sshpass installed, packet ssh and packet setup also type the instance password for you.
Every command,
on one screen.
packet <command> -h lists a command's flags.
| Command | What it does |
|---|---|
packet gpus | GPU types, regions and prices |
packet launch --gpu b200 [--region 7] [--name x] [--wait] | Launch a GPU; --wait returns when SSH is ready |
packet ps | Your instances and their state |
packet ssh <id> [--command cmd] [--copy] | Log in, run one command, or print the ssh line |
packet setup list · packet setup <preset> <id> | Install vscode, jupyter, jupyter-torch, workspace or full-dev on a running GPU |
packet logs <id> [--follow] | Instance status, refreshed until it ends |
packet note set <id> --note "run 4" | Label an instance so you remember what it was |
packet terminate <id> [--force] | Stop and delete; charges stop |
packet whoami · packet logout | Who you're logged in as; forget the key |
Everything the console does,
over HTTP.
JSON over HTTPS with a Bearer API key. Launches take an Idempotency-Key, so a retried request never starts a second GPU.
GET /instances/launch-options | Products and regions |
POST /instances | Launch |
GET /instances | List |
GET /instances/{id}/connection | SSH details |
DELETE /instances/{id} | Terminate |
# 1. pick a product and region curl -s https://dash.packet.ai/api/v2/instances/launch-options \ -H "Authorization: Bearer $PACKET_API_KEY" # 2. launch it curl -s -X POST https://dash.packet.ai/api/v2/instances \ -H "Authorization: Bearer $PACKET_API_KEY" \ -H "Idempotency-Key: $(uuidgen)" \ -d '{"product_id": "<id>", "region_id": 7, "name": "trainer"}' {"id": "cmr8x2k9a"} # 3. connect when it's ready curl -s https://dash.packet.ai/api/v2/instances/cmr8x2k9a/connection \ -H "Authorization: Bearer $PACKET_API_KEY" # 4. done curl -s -X DELETE https://dash.packet.ai/api/v2/instances/cmr8x2k9a \ -H "Authorization: Bearer $PACKET_API_KEY"
A GPU for one job,
then gone.
Launch a GPU at the start of a pipeline, run training or evals on it, and terminate it in a step that always runs. You pay for the minutes the job took.
- Works in GitHub Actions, GitLab CI, Buildkite or a Makefile
- API key from a CI secret, no interactive login
if: always()terminate, so a failed job never leaves a GPU running- Team budgets cap what pipelines can spend
name: train
on: workflow_dispatch
jobs:
train:
runs-on: ubuntu-latest
env:
PACKET_API_KEY: ${{ secrets.PACKET_API_KEY }}
steps:
- name: Install the packet CLI
run: |
curl -fLo packet https://dash.packet.ai/downloads/packet-v2.1.0-linux-amd64
chmod +x packet && sudo mv packet /usr/local/bin/
sudo apt-get install -y sshpass
- name: Launch a B200
run: |
ID=$(packet launch --gpu b200 --name "ci-$GITHUB_RUN_ID" --wait \
| awk '/Instance ID:/ {print $3}')
echo "ID=$ID" >> "$GITHUB_ENV"
- name: Train
run: |
packet ssh "$ID" --command \
"git clone --depth 1 https://github.com/$GITHUB_REPOSITORY app && cd app && python train.py"
- name: Terminate
if: always()
run: packet terminate "$ID" --forceLet your agent
rent its own GPU.
Coding agents already drive a terminal, so the CLI is all they need. For your own agent loop, wrap the API in a tool. Budgets and roles keep it inside the limits you set.
## GPUs When a task needs a GPU, use the packet CLI: - `packet gpus` to see types and prices - `packet launch --gpu <type> --wait` (note the Instance ID) - `packet ssh <id> --command "<cmd>"` to run work - Always `packet terminate <id> --force` when done. Prefer the cheapest GPU whose memory fits the model.
import os, uuid, requests API = "https://dash.packet.ai/api/v2" H = {"Authorization": f"Bearer {os.environ['PACKET_API_KEY']}"} def launch_gpu(product_id: str, region_id: int, name: str) -> str: """Start one GPU and return its instance id.""" r = requests.post(f"{API}/instances", json={ "product_id": product_id, "region_id": region_id, "name": name, }, headers={**H, "Idempotency-Key": str(uuid.uuid4())}) r.raise_for_status() return r.json()["id"] def terminate_gpu(instance_id: str) -> None: """Terminate the GPU so billing stops.""" requests.delete(f"{API}/instances/{instance_id}", headers=H).raise_for_status()
Your next run
starts at $0.39.
Sign up, top up, launch. You'll be on a GPU in minutes.