Introduction
If you've been following along here on Sandbox99 Chronicles, you already know the last few weeks have been busy. We moved the whole blog off WordPress and onto a self-built Python stack, I shipped a self-hosted YouTube Downloader, and I even wrote up a plain-English explainer on Claude Code Plugins (yes, the one with the Matrix kung-fu analogy). What tied all of that together behind the scenes was one tool: Claude Code, Anthropic's terminal-based AI coding agent.
I gave it exactly one month. A single subscription cycle. The goal wasn't to marry the tool — it was to actually learn it, ship real things with it, and form an honest opinion. Now that the month is up, I've decided not to renew right away. Not because I'm disappointed — quite the opposite — but because I'm a curious tinkerer, and I want to spend a cycle putting other AI assistants through the same wringer as my daily driver.
So consider this a friendly, no-hype field report: how I used it, what worked, what didn't, and everything I managed to build along the way.
How I Actually Used It
A quick note on context, because "is Claude Code any good?" is a meaningless question without knowing who's asking. My day-to-day wears a few hats:
- A bit of DevOps (containers, deployments, self-hosting)
- A little vibe coding (letting the AI drive while I steer)
- Some security analysis and homelab pentest
- And a fair amount of software architecture — planning before building
For most of the month, I lived in the terminal with Claude Code rather than any GUI. That detail matters for the pros and cons below, so keep it in mind.
The Pros and Cons (My Honest Take)
Everyone's mileage varies here, so this is my experience, not gospel.
👍 What I Liked
The subscription pricing felt fair for how I work. I'm not a heavy, all-day-every-day user, and the plan matched that reality without feeling like I was overpaying for capacity I'd never touch.
The usage limits are learnable, not scary. Claude Code works on a rolling 5-hour usage window, and that budget is shared across the models it reaches for during a session. There were days I hit 100% in under three hours — usually when I was deep in a big build. My advice? Don't sit there anxiously watching the meter. Just use the thing naturally. After a week or two you develop a feel for your own rhythm — when to go heavy, when to pace yourself — and the limit stops being a mystery.
👎 What Held Me Back
The polished desktop experience felt like a Mac/Windows perk. I run Linux Mint and Debian as my daily drivers, and for most of my month, the nice graphical apps — Claude Desktop and Claude Cowork — were things I read about but couldn't run. If you're on macOS or Windows, that all-in-one desktop window (Chat, Cowork, and Code side by side) is genuinely part of what your subscription buys, and I felt the gap living on the CLI.
⚡ Timely update: Right as I was wrapping this up, the picture changed. On June 30, 2026, Anthropic shipped an official Claude Desktop beta for Linux, installable straight from their own apt repository. It brings the same three-tab experience — Chat, Cowork, and Code — that Mac and Windows folks already had.
The fine print worth knowing:
- Officially tested on Ubuntu 22.04+ and Debian 12+ (x86_64 and arm64). Other Debian-based distros — Linux Mint included — may work but aren't formally tested yet.
- The full desktop experience is on all paid plans; the Code tab (which touches your local files) needs Pro, Max, Team, or Enterprise.
- Cowork on Linux runs inside a KVM-backed VM, so you'll want hardware virtualization enabled.
- Still missing on the Linux beta: Computer Use, voice dictation, and the global quick-entry hotkey on native Wayland (it works fine on X11).
- Fedora and RHEL aren't supported yet.
So the biggest "con" on my list is already on its way to being solved. Funny timing.
What I Shipped in One Month
Here's the part I'm genuinely proud of. This wasn't a month of "hello world" demos — Claude Code helped me put real, working software into the world.
🏆 The milestone: migrating Sandbox99 off WordPress onto a self-built Python stack. This was the big one — a full rebuild of the site I'm writing on right now, with a published result. Tech stack: Django/Flask, Jinja2, Bootstrap 5, and SQLite3.
Published Apps
- Network Architect Planner — a tool for sketching and planning network layouts.
- Secure Bookmark Manager — self-hosted, privacy-first bookmarking.
- YouTube Downloader — the one I've already written about here; a self-hosted, responsible-use downloader.
- Analitika — self-hosted, private web analytics (no third-party trackers phoning home).
Personal / Unpublished Builds
Not everything is meant for a public repo. A couple of these are just for me and the people around me:
- Offline YouTube Clone — I run this at home and share it with friends. Yes, there are a hundred similar projects on GitHub, but this one's mine, built on Python Django, Jinja2, Bootstrap 5, and Caddy.
- Android Offline Music Player (a WinAmp-style clone) — purely for personal use, and not headed to the Play Store. It pairs nicely with my YouTube Downloader, which handles pulling audio down as
.mp3. It's a native Android app built with Kotlin, a KVM emulator for testing, VS Code, and Claude Code. I'll do a proper write-up on this one soon — stay tuned.
My Most-Used Claude Code Plugins
If you read the plugins explainer, you know these are like Neo downloading a new skill in The Matrix. Here's my actual daily loadout — the ones I keep enabled:
| Plugin | What it does for me |
|---|---|
| caveman | Trims the chatter. Claude answers in terse, telegraphic "caveman-speak" — fewer filler words, same byte-exact code — which saves output tokens on long runs. |
| frontend-design (official) | Pushes the UI toward polished, intentional design instead of the generic "AI-looking" defaults. |
| modern-python (Trail of Bits) | Modern, security-minded Python practices — fitting, coming from a firm known for security work. |
| playground (official) | My scratchpad for quick experiments and trying ideas out. |
| skill-creator (official) | For building, refining, and measuring my own reusable skills. |
| superpowers (official) | The heavyweight. A full brainstorm → spec → test-driven-development → subagent-with-review workflow. It makes Claude slow down and plan before it codes. |
| understand-anything | My "explain this to me" button for unfamiliar code and concepts. |
If you're just starting out, I'd point you at superpowers and caveman first — one imposes discipline, the other keeps your token budget lean.
A Few Tips From My First Month
- Don't babysit the usage meter. Use it naturally and you'll learn your own limits faster than any dashboard will teach you.
- Keep a human in the loop. My favorite workflow is me as architect and approver, Claude Code as the builder. I review; it executes. That separation kept quality high and surprises low.
- Lean on plugins. They're the difference between a generic assistant and one tuned to how you work.
Final Thoughts
One month in, I'm a happy camper. Claude Code helped me ship a blog migration, four published apps, and two personal projects — and it genuinely taught me a lot about pairing with an AI agent day to day. That's exactly the experience I was after.
So why pause? Because that curiosity that made me try Claude Code in the first place is now nudging me toward the other assistants on the market. I want to run them through the same real-world gauntlet and report back — honestly, the way I've done here. This isn't a goodbye to Claude Code; think of it as a research sabbatical. And who knows — with an official Linux desktop now landing, I might be back sooner than I planned.
