Beanbox VM

Beanbox is a microVM built for LLMs and humans. The core idea is to allow LLMs to control a computer in isolation and in a safe way. Additionally, the interaction with the LLM should be like pair coding. You can write a few commands and the LLM can continue from there.

The design principles of Beanbox are:

• Simple
  • one command beanbox run
  • the interaction should be as simple as sharing a terminal with anyone else
• Fast: boot times should be in the 100-200ms region
• Secure: both you and your LLM can go rouge in the Beanbox without causing harm to the host machine

Demo

Run on your machine

I've so far only ran Beanbox on aarch64-apple-darwin but it should work on linux (x86_64) as well.

⚠️ Early days: expect rough edges

tl;dr

There's currently no official release so the way to run this is by building from source. To do this run:

brew install libkrun podman just

and then

just build

I wanna know more

Beanbox relies on the libkrun virtualisation framwork for creating VMs. On linux libkrun uses the KVM hypervisor and on macOS the Hypervisor.Framework. Libkrun is available on homebrew: brew install libkrun.

Next up you need to prepare a rootfs. You can in principle unpack any OCI container image and use that as the rootfs. There's a recipie in the justfile that relies on podman for unpacking the .tar and automatically sets the dns nameserver to 1.1.1. in resolv.conf. The default image used in ubuntu:latest. Once you have podman installed run just rootfs.

Run just build-gotty which builds the fork of gotty and places the binary in the rootfs under /bin/gotty. Gotty is launced on VM startup and is what facilitates the terminal sharing between the host and guest.

Finally just build-cli builds the cli which manages the Beanbox VM. On macOS it sets the correct codesign entitlements as well.

You should now have a binary called beanbox in the root of the repo. Happy hacking!

Usage

In browser

Once the Beanbox is running you should see a message similar to this (port may vary):

Open a browser on the assigned port and run commands in the Beanbox. Like so:

Example with LLM

For a more more smooth interaction, I've provided a simple app in /example that makes GPT-4 run code in Beanbox. This is the same code as in the demo video above.

To get your env set up run these:

python3 -m venv venv
source venv/bin/activate
pip3 install -r requirements.txt

We're also using a selenium chrome driver so you'll need to have Chrome installed. Alternatively you can swap it for something else.

Set your OpenAI api key and just run the main

export OPENAI_API_KEY="<YOUR KEY>"
python3 main.py

It opens the browser based terminal which is connected to the Beanbox microVM and should get a view like this in your terminal: Just go ahead and prompt the llm.

> what os is this?

Acknowledgments

This project is largely inspired by krunvm and build on top of libkrun. Gotty deserves also a special mention for the terminal sharing.