Skip to main content
1

Install

Node 22+, and a container engine for the sandbox — podman or docker. Check with npx nativesandbox doctor, or see requirements.
2

Run a command in a sandbox

The file you wrote was hydrated into a fresh sandbox, the command ran there, and the file the command created was synced back. The sandbox is disposable; the workspace is what survives.No model is needed for any of this.
3

Add a model

Reads ANTHROPIC_API_KEY. Four other adapters ship — OpenAI and compatible services, Gemini, Cohere, and local models through Ollama. See models.
4

Let it work

Tokens stream as the model writes them, commands print as they run, and the files it changed are in the workspace when the turn is done. Every event is persisted before it is emitted, so a client that reconnects can replay exactly what it missed.
5

Do it for a second user

A different workspace, a different sandbox, different files. Neither agent can reach the other’s work — see your users.

From the terminal

The same thing without writing any code:
See the CLI.

Next

How it works

What happens between chat() and the file appearing in your workspace.

Your UI

Rendering the event stream in a browser.