OpenShell: give AI agents a safe place to work

OpenShell gives AI agents a controlled workspace with clear rules for files, websites and access keys. Here’s how it works, what’s new and how to try a sandbox.

What it is

Imagine hiring an assistant to organise a folder of documents on your computer. You’d happily let them read that folder. But would you also give them your passwords, access to every other folder and permission to send files anywhere on the internet? Probably not. OpenShell lets you set those kinds of boundaries for AI assistants.

Some AI assistants can do much more than answer questions. These AI agents can open files, run software and connect to online services on your behalf. OpenShell provides a controlled workspace for them, called a sandbox. Think of it as giving the agent a room to work in, with specific doors it can and cannot open.

You decide which files, programmes and websites an agent can use through a set of rules called a policy. OpenShell checks those rules outside the AI agent itself. That matters: a request to “please don’t touch my private files” relies on the agent remembering and obeying you. A restriction enforced by the environment can block the action even when the agent tries it.

OpenShell isn’t a chatbot or a new AI model. It’s the security layer underneath tools such as Claude Code, OpenCode and Codex. It’s aimed at people running coding agents or other automation, especially when those agents need to work with real files and online accounts.

What’s new

The OpenShell 0.1.8 release, published on 10 November 2026, makes those boundaries more dependable.

Safer handling of access keys. Some online services need a secret token before they’ll accept a request. OpenShell can hold that token on the trusted side of the workspace and add it only when a request is going to an approved service. Version 0.1.8 requires more explicit rules for unusual places those tokens might be sent, such as custom request headers or web addresses.

Fewer shortcuts around file rules. A shortcut inside a folder can point somewhere else on the computer. This release prevents a sandbox from starting when a file-access rule follows certain shortcuts through a writable workspace. That closes a subtle route around the intended boundary.

More predictable behaviour when things go wrong. The release fixes issues involving sandbox cleanup, failed database connections and diagnostic logs. Those may sound like background details, but they matter when an agent is meant to keep running without creating new problems for its operator.

The project has also joined the Cloud Native Computing Foundation (CNCF), with its development moving towards the community-led ai-openshell organisation. Its main NVIDIA repository remains the place to follow code and releases during that transition.

What else is similar to it

E2B provides isolated cloud computers where AI-generated code can run. It’s handy when an application needs to start a temporary environment, execute code and collect the result. OpenShell puts more emphasis on controlling an agent’s ongoing access to files, networks and credentials, with rules enforced outside that agent.

Daytona also provides programmable environments for development and AI agents. It focuses on creating and managing those workspaces, with tools for files, commands and integrations. OpenShell’s distinctive concern is what an agent is allowed to reach once it’s running.

Infisical is a companion rather than a direct alternative. It helps teams manage sensitive credentials and permissions. OpenShell adds a separate layer by controlling which actions a running agent can actually take. Teams may use both.

You can also run an agent in an ordinary Docker container. Docker provides isolation, but OpenShell adds agent-focused controls for outbound connections, approved services and credentials. The best fit depends on whether you simply need somewhere to run code or you need to govern what an increasingly independent agent can do.

Why it’s different

OpenShell has one particularly useful rule: access has to be allowed, not merely left unblocked. Suppose an agent needs to read a project and download a package. You can allow the project folder and the required download service without opening your whole computer and every internet destination.

If the agent tries to contact another website, OpenShell can deny that connection and record what happened. An operator can review the request and decide whether to add a narrowly scoped rule. OpenShell even checks proposed rule changes for risky new permissions before they can be approved automatically.

The separation of responsibilities is the clever bit. The agent concentrates on its task. A trusted component outside the agent makes the permission decisions and handles credentials for approved destinations. It means an unexpected instruction inside a document or webpage cannot simply persuade the underlying security rules to disappear.

This is most useful when agents do real work: editing software, reaching business systems or running repeated tasks. For an occasional chat with an AI assistant, there’s little reason to install a whole sandbox system. But once an agent can use tools on your behalf, the question changes from “Can it do this?” to “What exactly should it be allowed to do?” OpenShell gives you a practical way to answer.

How to get started

  1. OpenShell is a tool for running agents, rather than an app you open and chat to. You’ll need a computer with a supported setup: Linux, a Mac with Apple silicon, or Windows using WSL 2 (experimental). You’ll also need a supported container or virtualisation system, such as Docker or Podman. The official installation guide walks through the options.
  2. For the first experiment, you don’t need an AI account or a paid model. We’ll start by creating an empty sandbox, just to see how a controlled workspace behaves.
  3. With your container system running, install the OpenShell command-line tool using the command in the official instructions: curl -LsSf https://raw.githubusercontent.com/NVIDIA/OpenShell/main/install.sh | sh. This installs the CLI and sets up a local gateway, which manages the sandboxes.
  1. Open a terminal and type openshell --help. If you see a list of commands, OpenShell is installed.
  2. Create your practice workspace with openshell sandbox create --name demo. The default image gives you a small Ubuntu environment with no AI agent installed. You should arrive at a terminal inside the sandbox.
  3. Try pwd or ls inside that sandbox. Now open a second terminal on your computer and run openshell policy get demo --full. You’ll see the rules that control what the sandbox can reach.
  4. When you’ve finished, leave the sandbox shell with exit, then remove your practice workspace from your computer’s terminal with openshell sandbox delete demo. Ready to try a real AI agent next? The official first-agent tutorial shows how to run OpenCode with an approved model provider and controlled access.
  1. The command isn’t recognised? Run openshell --help again after reopening your terminal. Check the installation instructions for your operating system if it still cannot be found.
  2. The sandbox won’t start? Make sure Docker, Podman or your chosen runtime is running. The support matrix explains which combinations work, and openshell sandbox list can show what was created.
  3. Something inside the sandbox cannot reach a website? That may be OpenShell doing its job. Network access is restricted until a policy allows it. The first network-policy tutorial demonstrates how to inspect a denied request and grant only the access you intend.
  4. Connecting a real agent asks for credentials? Use the official provider setup. OpenShell can store a credential separately and supply it only to an approved destination, instead of putting the secret directly into the agent’s workspace.

Explore further: OpenShell on GitHub · How OpenShell works · Run your first agent · Try a network policy · Release notes.

Share your thoughts…?

    Get Geekist in your inbox

      Choose your newsletters. Confirm your email to subscribe.

      Choose your newsletters

      Read the privacy policy