Thoughts on OpenCode

Harness

OpenCode is my harness of choice at the moment, both at work and on personal devices. It’s a pretty good zero-config harness. I use it a few different ways:

Locally on…
- My work laptop for all agentic work like coding, but also searching Webex, docs, our issue tracking system, and more.
- My personal laptop. I do most coding on a remote server, but I’ve used it here for things like improving the efficiency of my system (T14 gen2 running Omarchy)
- My desktop. One interesting thing I’ve done here is automating the process of ripping Blu-Rays, transcoding, and storing these on my NAS.
- My homelab server. I’ve used this for basically all setup tasks once the OS was installed.
Remotely over Tailscale on…
- My phone and other personal devices, connecting to the web application. This means I can vibe configure infra on my lunch break or on the couch at night!
- On my desktop I’ve been opting for the Desktop app (currently in beta). It offers the same functionality as the web app, but it’s easier to keep track of a desktop app than a browser tab.

OpenCode is pretty good at configuring itself - it has a built-in skill customize-opencode for editing its own config. You can just tell it to configure an MCP or plugin, and it will probably get 90% of the way there on its own.

One thing I’ve ran into is OAuth issues when adding new MCPs, because the auth flow expects the terminal (on homelab1) to be on the same host as the browser (my desktop or laptop), which it is not. This is easy enough to fix with SSH port forwarding, but it’s annoying when I have to do this. I’ve considered writing a custom plugin for this but haven’t gotten around to it.

Using opencode debug config is a useful check after making changes.

Speaking of plugins, there’s a lot! They also have an SDK.

I haven’t used their v2 release much, but it has some nice changes like a default password when running the web server.

OpenCode Go

Go is my provider of choice for my personal use. I don’t want to be locked into a single model family (Claude, GPT, GLM, etc.), but I obviously don’t want to pay API prices for models either! So for $10/mo you get, at minimum, $15 of usage for their models included in Go. There are limits in place for 5-hour / weekly / monthly chunks.

Some models have up to $60/mo in usage because of bulk discounts or reserved capacity. I definitely recommend using the models with $30 or $60 usage! Many of these (Deepseek V4.1 Flash, Muse Spark 1.3, GLM-5.3-Flash) perform nearly as well as the lower use models, and you get to use these MUCH more than you’d otherwise get to for $10/mo.

Closing

There’s a lot of harnesses and it’s a bit overwhelming trying to learn and compare new ones. I continue to use OpenCode because it’s what I’m used to. I’m also playing with Pi and Hermes a bit, but they haven’t stuck (yet). Go continues to offer tremendous value for my personal work, and there isn’t a great alternative available elsewhere.