It’s no secret right now that AI is everywhere, and it’s expensive. Trillions on trillions of dollars have been poured into the medium over the last decade. Providers recognize this, as one by one they all fall like dominoes transitioning into usage-based pricing models. This places additional pressure on the customer to make every single token count, as hundreds, if not thousands of dollars are at stake throughout AI-assisted project development.

But what if it didn’t have to be that way. What if you could wield the power of trustworthy, capable generative AI models from your own home, nay, your own computer? Imagine the leg-up that would give small businesses without the tens of thousands of dollars to spare on cloud compute? Imagine what you could create in your very own home!

I’ve been able to accomplish this thanks to the help of the open source community, and in my first post on this blog I’d like to explain why that’s relevant to you, and why I think it’s worth checking out if you’ve got the hardware.

This is not a tutorial, there are plenty of those online to walk you through the process of setting this up if you’d like to try it for yourself, but there are two pieces of software that make up the foundation of this system. The first is LMStudio, a desktop application that lets you download, run, and experiment with open-source LLM’s (Large language models) on your own hardware. The second, ZooCode. ZooCode is a plugin (Forked from the now deprecated RooCode, RIP) for Visual Studio Code that essentially orchestrates an agentic system around your coding interface. You can provide credentials from a loaded model in LMStudio within the ZooCode plugin, utilize that model in ZooCode’s agentic framework, and it can support your development endeavors from directly within the Visual Studio Code coding interface.

Now the hardware required to do this is beefy but not crazy anymore. Model recommendations vary based on hardware, as you’ll want the model to be contained entirely within your VRAM, which is essentially dedicated video memory supplied by your systems GPU (Graphics Card). My system currently utilizes an NVIDIA RTX 4070 with 12GB of VRAM, which is probably closer to the low end of the kind of hardware you’d want to have to use this kind of system, although it’s serviceable if you have similar. There are plenty of articles out there discussing which models are best for specific hardware setups, so do some window shopping until you pick one out that you like (I really like “qwen3.6-35b-a3b-mtp [Quant: IQ4-XS]” right now).

ZooCode can be configured to read and write to files within the set project directory, and has configuration settings allowing the user to govern exactly how much autonomy and control the plugin has over files in the set workspace. Because ZooCode is a Visual Studio Code Plugin, it lets the model read and write within the set space you’ve allowed it to edit on your local drive. That means that the data, tokens, and output, every stage of the AI Life Cycle, takes place ONLY on your device.

That’s right, PRIVATE Generative AI!! That’s nuts, right?

And the best part is – it’s pretty much free! Everything is maintained by the open-source community and therefore, there’s no direct access cost to any of this. However, the “Free” I placed in the title might be hyperbole. Hardware costs have ballooned due to the expansion of this kind of technology, so the upfront cost to get into it if you’re not already a massive gaming head like myself can be a significant blocker for a lot of people. However the cost outside of that is extraordinarily minimal. The electricity cost is mere cents, as these open-source models are extremely efficient. In fact, it’s more expensive to run my microwave for an hour, than it is for my PC to run my preferred model for that same period of time. Additionally, there’s no usage limit, so your chosen model can help you as much as you’d like without having to worry about being cut off by big AI!

As mentioned in the title, ZooCode operates as an agentic system. There are multiple “Modes” (Agents) that are essentially configured personas that determine how the underlying model behaves. These don’t quite meet the definition of an agentic system by themselves, but the most powerful mode, Orchestrator, is what brings the whole thing together. Orchestrator decomposes the set task and delegates subtasks to other mode-instances, coordinating them in the familiar orchestrator/sub-agent pattern you usually see in agentic systems like this. The user is free to prompt any one of the modes individually if there’s a task one of them does best, or for beefier tasks or entire projects, lock in with the Orchestrator mode as it deconstructs the original ask into sub tasks and automatically swaps to the appropriate mode to accomplish them. When Orchestrator swaps to other modes whilst performing a task, it gives that mode a set of instructions on what to do to accomplish the provided sub task, so in essence, each mode given a sub task has its own context window for the purpose of accomplishing that sub task.

Now why even bother with this if the results aren’t good. Well, what you’re measuring is the result of the model itself and how well you orchestrate that model through the framework Zoo provides, so it wouldn’t be fair to just list a bunch of stuff I’ve built with it and say it’s great. You might have better or worse hardware than me or be more limited/more free in the models of choice you have. But I will say this, the very blog you’re reading this on was entirely configured and setup by my ZooCode Qwen assistant, and the net electricity cost to finish it was likely well under a single dollar.

As AI continues to dominate this space, there’s room for individuals with hard skills to configure and adapt to the ever-changing environment it’s brought upon us. I believe ZooCode is an exact example of that adaptation, and I’ve used it myself to create some pretty cool stuff, or just to have fun playing around with a cool Github repo for a project I’m interested in. It’s entirely on-device so I’m not risking my data being sent elsewhere, its capable enough to perform some pretty impressive tasks given its local, all orchestrated through ZooCode’s agentic framework and I don’t have to send my money to a billion dollar corporation in order to do it.

So yeah, ZooCode is pretty cool. If you do check it out, let me know your experience, or, feel free to share anything cool you build with it! Looking forward to the discussion. -C