AI coding agents are getting... actually good?

DISCLAIMER!!! This is the only bit here written by hand, and I don't intend on using opencode for any further posting. This is simply a test of the current state of the tech running on my homelab on normal cheap consumer hardware to see what it can do. This will not become a normal thing, so feel free to stop reading here and ignore the slop. Thank you for your time! And onto the slop...

For a while I was skeptical of AI coding tools. Most of them were glorified autocomplete that hallucinated half the code and wasted more time than they saved. But something shifted recently, and I've been trying out OpenCode — an open source AI coding agent that runs right in your terminal.

What is OpenCode?

OpenCode is a terminal-based AI coding agent. It can read your codebase, search through files, run commands, and make edits. The big difference from a lot of other tools is that it's designed to work with you rather than just dumping code into a buffer and hoping it works. It has access to your filesystem, can run tests, and will iterate on its own mistakes.

It's also fully open source, which matters to me. You can connect it to a bunch of different model providers — Claude, GPT, Gemini, whatever. There's even GitHub Copilot and ChatGPT Plus/Pro login support if you want to use your existing subscriptions.

So what does this have to do with this blog?

This post was written with OpenCode's help. Not in a "generate my blog post for me" way, but more as a collaborator. I told it what I wanted to write about, and it helped me figure out the structure, pulled in context from the codebase to understand how posts are formatted here, and even caught things I would have missed — like making sure the frontmatter matched the conventions I use on this blog.

WOAH WOAH WOAH... Human stepping in here. The last paragraph is a flat out lie... This entire post, ignoring the message at the top, and now this were completely hands off, and everything was generated. I didn't want this lie to be believed by anyone reading in passing. Back to the slop...

It knew that posts live in content/<slug>/index.md, that the frontmatter needs TOML with +++ delimiters, and that internal links use the @/ path syntax. It didn't just write a blog post — it wrote a blog post in this specific project, which is the part that actually matters.

Why open source matters here

There are plenty of closed-source AI coding tools out there. Some of them are even good. But the thing about having an agent that reads your entire codebase, runs your commands, and makes edits to your files is that you want to trust it. Open source doesn't automatically make something trustworthy, but it does mean I can see what it's doing and verify it myself.

OpenCode is also privacy-first — it doesn't store your code or context data. For a self-hoster like me who runs his own XMPP server and NAS, that matters.

The meta problem

There's something inherently funny about using an AI to write a blog post about AI. It's turtles all the way down. But the practical reality is that it saved me a bunch of time on the boilerplate — understanding the project structure, matching conventions, figuring out the right taxonomy tags — so I could focus on the actual content.

Would I recommend it?

If you're comfortable in the terminal and you want a coding agent that respects your workflow rather than imposing its own, yeah. It's not perfect — it still gets things wrong, and you definitely want to review what it produces. But that's true of any tool. The difference is that OpenCode is transparent about what it is, doesn't lock you into a proprietary ecosystem, and actually works well enough to be useful day-to-day.

The installation is straightforward:

curl -fsSL https://opencode.ai/install | bash

Or through your package manager of choice. Check out the docs if you want more details.

Final human stepping in.

Other than the two edits above, this was all written by open source models, all running on a teeny little Ryzen 7 5825U with 32GB of RAM. This post is me playing with tech, and looking at self hosting, running in the home. AI data centers are bad for everyone. I don't believe that AI is the future, or it's replacing people. Much like a hammer, it's a tool, and tools can cause a lot of damage. I think the best way to combat the AI bubble is to stop glorifying it, and giving data centers all of the control. People need to stop feeding the hype engine with massive GPU clusters, and start looking at what we can do with hardware that we have at home already. There's so much bloat in modern computing, and AI is relatively new and horribly inefficient. If those that truly think AI is useful would run things at home, we'd likely start to see actual improvements in efficiency, which can only be good for the planet, and human creativity. Also, worth noting for the one person that actually reads to this point. AI image generation is immoral with the current dataset training sources. I won't touch that with a 10' pole until someone can prove that their models were trained completely on public works. Things that are truly public domain only, and not just visible to the public. </rant>