It used to be true that you have to use a sota model to get the best results. But for people who are in the tech space, they know that small models have an improvement trajectory that's catching up incredibly fast. Soon the small models will be able to perform most of the same tasks that you need on a daily basis.

But even if you want to use a local model there's not a great harness that provides all of the capabilities that you would want out of a comprehensive agent. Many of the harnesses are currently locked down or built for API keys, not optimized for the local models. This is why I built Capy, short for capybara. It leverages other open source projects (openCode and Ollama) to create the same codex experience that you're used to, but with the confidence that everything stays local.

As an engineer, the I focused the first use cases on the coding experience, but with local models, it opens the door for processing all of your sensitive data without worrying about what the labs will do with it. With gemini, I would never share any tax forms with it, but with a fully local agent, that concern evaporates.

I'm curious what other use cases people would find valuable from a local agent that they wouldn't use with regular agents controlled by large labs?