If you caught my earlier post evaluating OpenCode against Claude Code, you know I liked what I saw on paper — privacy posture, subscription flexibility, bring-your-own-key support. But "liked on paper" and "trust it with real work for 30 days" are two different things.

So I did the 30 days. I ran it as my daily driver AI harness alongside Claude Code, tested the free tier, burned through a month of paid credits, and threw a personal project at it to see how it held up. Here's the unfiltered rundown.

First Impressions Still Hold Up

A month in, the interface still feels clean and easy to navigate. No hunting through nested menus to find a setting, no clutter competing for your attention. If you've bounced off other AI CLI tools because the UX felt like an afterthought, OpenCode doesn't have that problem.

Usage Monitoring: Familiar Territory

If you're coming from Claude Code, the usage tracking model will feel immediately familiar — OpenCode monitors usage across 5-hour, 7-day, and 30-day windows, the same rolling-window approach Anthropic uses. If you've already built a mental model for pacing your Claude Code sessions, that instinct transfers directly here.

The Free Tier: Great for Experiments, Not for Secrets

The free tier is genuinely useful for tinkering and personal projects — it's a low-friction way to kick the tires without pulling out a credit card. But here's the catch I want to flag clearly for anyone reading this before diving in:

Be careful what you feed a free tier. Free-tier usage is commonly used to help train the underlying models, and this isn't unique to OpenCode — it's standard practice across most AI providers offering a no-cost option. If you're prototyping something with proprietary code, client data, or anything under NDA, the free tier is not the place to do it.

Treat the free tier the way you'd treat a public Pastebin: fine for throwaway experiments, not fine for anything sensitive.

Subscription Tier: One Month, Every Model I Could Reach

I committed to a single month of the paid subscription specifically to stress-test as many available LLMs as I could against my pet project. This is where OpenCode's model flexibility really shows — instead of being locked into one vendor's model lineup, you get to shop around and see which model actually fits your workflow.

Bring Your Own Key (BYOK) Actually Works

If you'd rather use your own API keys from OpenAI, Claude, Gemini, or other providers, OpenCode supports that too. I tested this personally with my own OpenAI key, and it worked without any friction — no weird workarounds, no broken integrations. If you're already paying for API access elsewhere and don't want a second subscription, BYOK is the way to go.

Pay-As-You-Go: Where the Free Tier Becomes a Safety Net

OpenCode also offers a pay-as-you-go model for when you don't want a recurring subscription. This is where the free tier earns its keep: if you run out of subscription time or PAYG credits mid-task, you can fall back to the available free-tier models instead of getting hard-stopped. It's a nice bit of continuity that keeps you working instead of staring at a paywall.

No Subscription? No Problem

Here's a detail worth calling out for anyone hesitant to commit: you don't need to subscribe or load PAYG credits just to try the tool. After installation, you can still use it — no gatekeeping just to kick the tires.

Zero Data Retention: Real, But Read the Fine Print

This is the part that matters most if you're doing anything remotely confidential — and it's also more nuanced than I first assumed, so let me correct myself here.

OpenCode's own paid offering, Zen, does carry a zero-retention commitment for its directly-provided models, and OpenCode's terms of service state they won't retain your content when you're using Zen. That's the good news.

The catch: if you're using third-party models through OpenCode — including via BYOK — your content is subject to that provider's retention policy, not OpenCode's. OpenCode is upfront that it doesn't control those providers' practices. And even within Zen itself, there are named exceptions: requests routed through OpenAI or Anthropic APIs are retained for 30 days per those providers' own policies, and a handful of free/trial models (like Big Pickle and a few others) may have their data used to improve the model during the free period.

So the takeaway isn't "everything paid is zero-retention" — it's "check which model you're actually routing through." If PII or NDA-bound work is on the table, stick to Zen's models that explicitly carry zero-retention, and don't assume BYOK automatically inherits OpenCode's privacy posture — it inherits your provider's instead.

The Verdict: My Backup Daily Driver

After a full month, OpenCode has earned a permanent spot in my toolbox — not as my primary AI harness, but as a solid, privacy-conscious backup when I want model flexibility or need to step outside my usual Claude Code workflow. If you're already running an AI-assisted DevOps setup and want a second option in your back pocket, OpenCode is worth the month it takes to find out for yourself.

Have you tried OpenCode or another AI harness tool as a Claude Code alternative? I'd love to hear how it held up in your workflow — drop a comment or reach out.