I let an AI coding agent rebuild my entire website — a hands-on ZCode review
Disclosure, because this site doesn't do hidden stuff: the ZCode links below point to the official site — plain links, no referral code attached yet. If that ever changes to an affiliate arrangement, this paragraph is where I'll say so. Everything below is what actually happened, including the parts that didn't go smoothly.
This website is a fraud, in the nicest possible way. It looks like a hand-built retro terminal, and it is — but the hands were partly an AI coding agent's. Every page you've read, the speedrun game, the leaderboard, the contact inbox, the RSS feed, even this article's publishing pipeline — all of it was built and deployed with ZCode, z.ai's terminal coding agent. So this isn't a review from a weekend demo. It's a review from a production site I run, with subscribers, backups, and a nightly cron attached.
What ZCode actually is
ZCode is a coding agent that lives in your terminal, built by z.ai around their GLM models. You describe what you want in plain English; it reads your codebase, plans, edits files, runs commands, and reports back. The GLM Coding Plan (from around $18/month at the time of writing — z.ai says the same subscription also works in 20+ other coding agents like Claude Code and Cline, so you're marrying the plan, not the tool).
If you've seen Claude Code, the shape is familiar: terminal-first, agent-style, tools included. The differences that mattered to me were price and where the plan works — not benchmark flags.
What it actually did on this site
Concrete outputs, so you can judge the tool by its artifacts:
- The whole terminal UI — a zero-dependency CRT interface with themes, a markdown renderer,
tab-completion, and a blinking caret at a suspiciously precise 1.06s. No frameworks, matching the site's no-npm philosophy.
- The speedrun game — timers, a SQLite leaderboard, anti-cheat (it rejected my own
machine-speed test run as "not humanly possible," which I found both insulting and correct).
- The boring critical stuff — nightly automated database backups, a rate-limit fix, an
XSS-safe leaderboard, path-traversal hardening. The unglamorous work nobody tweets about.
- This publishing system — a content bot on a daily cron that drafts social posts and essay
outlines into a folder for me to review. The bot never posts; it drafts. That boundary was one sentence in a config file, and the agent respected it for weeks.
The workflow that worked best: I describe a change in a sentence, the agent proposes a plan, I correct the plan, then it executes across a dozen files while I drink coffee. The plans are the product — I've caught more mistakes in the two-minute plan review than in any code review.
Where it stumbled (the part most reviews skip)
- Terminal-only. There's no GUI. If the command line frightens you, this whole category
will — my own site needed an /easy page for non-terminal people, and that irony is not lost on me.
- Plan-review is not optional. Hand it a vague goal and it will confidently build you a
cathedral where you asked for a shed. The tool is only as good as the description you give it — which is less a bug and more the actual skill you're buying.
- First-run friction exists. Community reports of setup quirks on some machines match a
bump or two I hit early on. Nothing that survived a restart, but it wasn't magic out of the box.
- You must keep the leash. Left fully unsupervised on a deploy, an agent will do exactly
what you said rather than what you meant. The deploys went fine because I read the plan.
The money part
Here's the economics that made me keep the subscription: a full site rebuild — design, backend, database, deployment, plus ongoing feature work and this content pipeline — ran inside a plan that costs less than a few hours of a freelance developer. Your mileage is your mileage, but for a founder who ships, the math wasn't close.
If you want to try it: the official site is zcode.z.ai, and the GLM Coding Plan works across the major coding agents. Using a referral link takes a discount off your first subscription — this article doesn't have one attached yet, so those links are plain and untracked.
Who it's for
If you can use a terminal and you ship things — sites, tools, internal automation — an agent like this is the highest-leverage subscription in the stack. If you want a GUI and a magic wand, wait a year; this category is moving fast enough that patience is also a strategy.
The site you're reading is the proof of work. You just finished the affiliate disclosure, and you're still here — that's the review.
This fits the bigger picture of running AI on your own hardware and choosing the right tool for each job — an AI coding agent is just one more tool that has to earn its place.