HarnessVM

HarnessVM launches secure, double-isolated AI coding sandboxes with browser and terminal access, so you can run agents, tools, and local or cloud models in YOLO mode with confidence.

Comment 1

Inspiration

Agentic development is a big game changer. We can now run several agents in parallel, letting them work for minutes or even hours to translate product requirements into high-quality implementations. They can take care of planning, coding, and even testing the project completely on their own. For that, they need an environment to run and test things safely and freely, isolated from your computer and sensitive files.

What it does

HarnessVM lets you quickly launch secure AI sandboxes with graphical and terminal interfaces.

Simply go to your project's directory and run HarnessVM. It then uses Lima and Docker to create a virtual machine with a double isolation layer to ensure your agents run safely and freely. You can then jump into coding and interact with the sandbox, right from your browser, locally or remotely.

It comes preconfigured with VS Code, a browser, and your favorite coding tools and agents like Codex. It isolates your file system and private networks from the development environment.

With HarnessVM, you can unlock the full potential of an agentic development workflow and run in YOLO mode with peace of mind. You can also use HarnessVM to try different agent harnesses and on-device LLMs running on your host machine or in the cloud.

How we built it

Everything was done with the help of Codex, from research to planning, coding, testing, and enjoying!

Codex's generous subscription plan, paired with the frontier GPT-5.6 Sol model, made development doable in the matter of a couple of days. Of course, my experience as a senior software engineer helped to plan and guide the agent, but most of the heavy lifting was done by Codex.

Challenges we ran into

The biggest challenge was to avoid complexity. It's definitely easier said than done. What I found most helpful was the fact that GPT-5.6 Sol closely followed my instructions to think deeply about each change, do it in as few lines of code as possible, and minimize cognitive complexity. This makes a huge difference, because the end result is something other humans can easily understand and contribute to.

Accomplishments that we're proud of

With Codex, I got to implement an excellent prototype that I'm actually using for my other projects. It works smoothly and uses industry-standard, battle-tested dependencies like Lima and Docker. All you need to do is install the dependencies, run the harnessvm command, and you're ready to go.

What we learned

I've been working with AI and agentic coding tools frequently for 2-3 years now. What I can definitely tell is that models like GPT-5.6 Sol, especially when set to high thinking effort levels, can help save countless hours that previously would have taken days or weeks. They also follow instructions much better than before. Sure, they need supervision, but the productivity gains are real, and I can spend time on the higher-level, creative tasks.

What's next for HarnessVM

I've started testing it on my own projects, sometimes in parallel, thanks to the isolated machines that launch in their own container. I'm also planning to use HarnessVM to test various cloud and on-device LLMs to understand their capabilities, without having to worry about what could go wrong.

Built With

Share this project:

Updates