
When you use a tool every day for 3 months, you stop seeing its limitations. They’re there, of course. You work around them, adapt to them… and they become almost invisible.
Then comes the day you hit your Max x20 subscription usage limit… and the motivation amplified by necessity invites you to try a different interface.
Or rather, invites me -- because that’s the story of this week.
Built by the team behind Serverless Stack, OpenCode has become the most credible open source alternative for running LLMs from a terminal. Compatible with multiple models, equipped with a richer interface and a real-time sidebar, it deserves a closer look.
Here’s a tour of the main differences.
What Claude Code doesn’t do (and OpenCode does very well)… or the reverse
The usage sidebar: simple, but essential
OpenCode permanently displays in its sidebar:
- the current directory
- the session title (which you can change)
- token consumption (volume, percentage, and cost in $)
This is information I didn’t have directly in front of me in Claude Code, yet it proves invaluable for interrupting when the agent is going in circles or monitoring request costs in real time.
There is a workaround: this post by Freek explains how to add a custom status line to Claude Code, my own adaptation, or ccusage which I just installed.
But OpenCode integrates it natively, cleanly. For anyone looking to optimize sessions or simply keep an eye on costs, it’s a real game changer.
Plan / Build modes
OpenCode visually distinguishes two states:
- Plan: the AI thinks, proposes, questions.
- Build: it executes.
You switch between them with Tab. It’s a UI detail, but it imposes mental clarity: am I exploring or executing?
Claude offers similar functionality. You can activate plan mode via the /plan command, or simply specify in the request that you want a plan before execution. And in practice, this habit becomes natural.
The terminal
OpenCode works best in a modern terminal like Ghostty.
In VS Code’s integrated terminal or Ubuntu’s default terminal, some features are missing or degraded.
This is currently my main friction point. I work mostly from Codium and haven’t yet found an elegant solution to replace its integrated terminal with Ghostty. It’s probably possible, but it requires a configuration investment I haven’t taken the time to make yet.
Two ecosystems, two skill sets
The real underlying issue, which neither OpenCode nor Claude fully solves, is workflow portability.
I have a .claude folder versioned on Git: skills, context rules, reusable prompts. Everything is designed for Claude.
Even when another model is accessible via OpenCode (Kimi, for example), it doesn’t interpret these elements the same way. Kimi’s Chrome MCP doesn’t work like Claude’s dev-browser skill.
Unifying the interface isn’t enough if the underlying models speak different languages.
This portability problem eventually led me to run Claude Code and Codex on the same projects, with a single source for instructions and skills.
What other users report
Points I haven’t personally experienced in a significant way, but that come up regularly:
- Diffs are clear
- Permission management is predictable
- The cognitive load of using the tool should be near zero so you can focus on what you produce with it
And you, what’s your take if you’ve tried both?
Verdict
Short article, simple conclusion: there are no major differences (yet) between the two contenders.
OpenCode is a serious alternative. If Anthropic didn’t restrict its subscription to its own tool, I’d probably switch to OpenCode for the convenience of changing models easily and centralizing my skills in a single setup.
In the meantime, Claude Code remains my primary tool (particularly thanks to Opus). OpenCode stays within terminal reach.
This article will probably be obsolete in one to two months.
AI, SEO and Web strategy session
Is your team choosing between coding agents?
Compare workflows, cost, security, and migration effort using your actual environment and constraints.
€350 excl. VAT, up to 90 minutes
If you’d like to follow the evolution of my stack, you can check out the article My AI adoption journey or subscribe:
Continue reading: