Building Ai Agents Open Code And Open Source A Conversation With Dax

Building AI agents, open code, and open source: A conversation with Dax Announcing our Series F . Learn more

Community

Building AI agents, open code, and open source: A conversation with Dax

Dax Raad, creator of OpenCode and Zen, discusses the launch of Zen, the philosophy behind building in open source, and why terminal-based workflows matter.

Authors

Madison Kanna

Last updated October 9, 2025

Share

TL;DR In this interview, Dax Raad, creator of OpenCode and Zen, we discuss why Dax built OpenCode, the launch of Zen, the philosophy behind building in open source and why benchmarks often mislead users about actual product quality.

Q: What is OpenCode, and why did you start building it? Dax: We all started to use LLMs for our work. And it was always a very clunky experience. It was, I'm in my editor, I'm in the zone, I got all my key bindings. I'm really happy with how it works. I have a question, okay, I need to stop, switch over to my browser, type out my question, paste in some code, potentially submit, wait for the response, have a conversation. Then maybe I get some code that it spits out and I have to copy paste back into my editor. And this just felt really bad. I'm in  this loop that I'm gonna use constantly throughout the day, but it's not the right approach. And I kept thinking, browsers can technically access your file system. So I was just waiting for one of these LLMs to let you say oh, you can just mount a folder, which is your project, and you can mention files in your project. It can write results back into your files and imagine it's all in the browser. And I just thought, okay, someone's gonna do this one day. But then Claude Code came out where it ran in your terminal, alongside your editor. You can ask it stuff and it can just manipulate your file system directly. And it elegantly solved the problem I had, which was the LLM can't touch my files. And that was the first AI coding product that really clicked for me. I had the aha moment. And I started to use it. I tried a bunch of AI products in the past year, but that was the first one that stuck. Then I started to think, there's other models out there that I want to try. Every week there's a new cool model that's claiming X, Y, Z things. Also we have Terminal, our coffee shop. So I had a deep understanding of how much you can do in the terminal. And we had another project that we were working on that also had some really advanced terminal functionality. 'm also a Neovim user, so I kind of understand what the ceiling of what you can do in the terminal is. The scope that Claude Code defined makes a lot of sense. They're not trying to go crazy with terminal stuff. They just want something that's simple and works everywhere and friendly to the average person, but they're never gonna go and push the limits of what's possible in a terminal. And I felt myself wanting to do that. Claude Code's good, but again, there's certain things that I felt could be done better. We  started exploring this and thought, what if we built an open source version of Claude Code that was heavy on the terminal stuff. And also was really flexible for any model because that's where open source has an advantage. Q: Why make OpenCode open source? Dax: Just because something's open source doesn't mean it's going to be any better than the closed source equivalent. It only happens when there's a long tail of things to cover. We need to cover all the different models and people playing with all different models. We need help improving how it all works. Then it kind of makes sense to be open source so the community can help with that. And we're obviously very good at bootstrapping the open source community overnight, so it felt like the right problem for us. Claude Code, for example, is not open source. And I think that makes total sense because there's no problem that Claude Code has that a large open source effort would solve, right? Claude Code, it just works with Anthropic models, it's all tightly integrated. You don't really need help from the community too much. I'm sure they can fix bugs here and there, but the core reason for existing doesn't benefit from the open source stuff. On our side we need to make sure our stuff works really well across all the models. And new models come out every day. I can't go and use every single model every day. I'm not gonna know the quirks between them. I never use Gemini. This is not something I use. We're never going to deliver a good Gemini experience because we don't use it personally. But this is where the open source community can help because someone out there uses Gemini every single day and they're oh, Gemini makes these common mistakes when editing tool calls. Here's a fix for that. Or here's a bug with your API. We can fix that. So if you need large manpower for something you're doing, but you can't directly fund that effort, being open source helps quite a bit. And that's our core positioning. We are the open source option in this space. Q: What is an AI agent, really? Dax: I made so many jokes about this a while ago where I said, we just added a new word for no reason. But to be fair, I see the reason for the term. When you're using an LLM and you're just prompting it and getting a response, that's one way of using it. When you're giving it a goal and giving it access to tools and it's just going to keep running until it achieves that goal, that's a kind of different way of using it, and that's what people categorize as an agent. At the end of the day, it's just an LLM that has access to some tools. I can edit a file or I can read a file, or I can look up something on the internet. And it's going to think about what tool to call. It's gonna call tools, then it's gonna use that result as a new input. It's gonna think again, it's gonna call a tool, the result as a new input. It's gonna keep repeating that until it decides the task is complete. Or it doesn’t know how to solve this. And then you get the response back. Q: Tell us about Zen. What is it? Dax: So if you think about how I described OpenCode, it’s very flexible, very configurable. You can use whatever model, you can hack it, you can customize it. There's so...