Transcription
This week we've taken a look at Open Code. Open Code is a free and open-source AI coding agent. It's very easy to install with a oneliner command and it's highly portable. That means I can go ahead and build a Docker file for it and run it with Docker Run. And I really like the look and feel.
The coolest thing about Open Code is the low barrier to entry. You don't need an account, a subscription, no payment methods required. It has a connect command making it provider agnostic. So you can go ahead and hook it up to your existing OpenAI, GitHub Copilot, Anthropic Claude, or Google subscription. It has a ton of other providers. So you can hook it up with existing cloud models like AWS and Azure models. You can even run models locally like Gemma 4.
But the coolest thing I found is if you type models, it has a few free models pre-selected. So you can immediately start using it and start learning about the AI concepts. In our AI command line series, we learn about all the concepts. So we learn about agents, the default agents that come with Open Code as well as our own custom agent and how to build sub-agents. So we create a Open Code directory with an agents folder inside and then we build our agents.md. Here I have a technical writer agent that helps us plan, create content guides, as well as production video scripts, and also helps us spin up and provision test local Kubernetes clusters.
We learn how to define agent definition files. But the main important thing is to learn about context engineering. Context engineering starts with the init command where we generate an agents.mmd that gives our main agent more focus when working with our repository. Then we take it one step further by looking at sub-agents. So the agents command builds our own agent, and agents have specific context and specific focus. They can point to their own models, their own tools, and have their own custom system prompts.
We take context engineering even further by having skills defined. Open Code supports agent skills. So you can go ahead and define your own skills. Here you can see I have an agents folder with a skills folder. And I have four skills. One skill is specifically designed to provision local Kubernetes clusters. You can see that in the skill.md file. We have a skill name and description. It's important to know that only this metadata gets loaded into context when our agent starts up, not the actual body of the skill. And this helps us optimize the context even further. Only when we ask the agent to provision a Kubernetes cluster will it go ahead and load the entire skill. And this is the whole benefit of a skill.
Our skill comes packaged with pre-packaged and pre-built scripts. This means the model doesn't have to guess. So the skill gets a consistent, repeatable outcome every time, and all of these scripts don't enter the context window. So it helps us to further optimize cost.
And lastly, we also take a look at /mcp. And here you can see I have a Kubernetes MCP connected. MCP servers can either be local binaries like this one, or you can host it behind an ingress or gateway API as a web server.
Now, if you're interested in Open Code or any of the other CLIs we've covered on the channel, check out the link down below.