Skip to content

Sushi OpenShell

Sushi OpenShell packages NVIDIA OpenShell, an open-source runtime that runs commands and AI agent workloads inside sandboxes governed by explicit policies. The image runs on Ubuntu 24.04 for Graviton (ARM64), with Docker as the sandbox driver and a private gateway that only listens on the instance itself, secured with mutual TLS.

Launch on AWS Marketplace Getting started

What you get

  • NVIDIA OpenShell v0.1.2: the openshell CLI, gateway and sandbox runtime.
  • Docker Engine as the sandbox driver, with the OpenShell sandbox, supervisor and default workload images already loaded.
  • A local gateway on 127.0.0.1:17670, protected by mutual TLS. Each instance creates its own TLS identity on first boot.
  • A first-boot service, sushi-openshell, that configures and starts the gateway for the ubuntu user. Upstream telemetry is turned off.
  • A walkthrough script, /opt/sushi/docs/demo.sh, that checks the gateway and runs a sandbox end to end.
  • Host hardening: key-only SSH with root login off, automatic security updates, kernel hardening, and the AWS Systems Manager agent for SSH-free access.

No TLS identity, passwords or provider credentials are baked into the image.

You do not need a GPU, a model-provider key or an inference service to start the gateway and run sandboxes. This is a CPU image: it does not include GPU/CUDA, inference models or a MicroVM configuration. To use an AI model, you connect your own agent and provider after launch.

Ports

Port Protocol Used for
22 TCP SSH, as the ubuntu user. Optional if you use Session Manager.

The OpenShell gateway on port 17670 listens on 127.0.0.1 only and is never reachable from the network. Do not expose it.

A t4g.medium (Graviton, 2 vCPU, 4 GiB RAM), with a root volume of at least 32 GiB. Size CPU, memory and disk up for real sandbox workloads.

Billing

  • The software costs USD 0.10 per running instance-hour.
  • An optional seven-day free trial covers the software for one instance at a time. When it ends it turns into the paid hourly offer automatically. The installed software is the same during and after the trial.
  • Additional instances running at the same time are billed at the hourly rate.
  • AWS bills EC2, EBS and data transfer separately, including during the trial. Any model provider you connect bills its own usage.

Where to go next

Support

For help with the image and setup, email info@softwaresushi.com. NVIDIA support and model-provider credits are not included.

Upstream project

OpenShell is developed by NVIDIA and released under the Apache-2.0 license. Software Sushi packages it for AWS and does not claim NVIDIA authorship, endorsement or support. See the NVIDIA OpenShell repository and the v0.1.2 release.