← sources

3 years to build "one repo → one URL". Claude Code made the code fast. The product still took months.

source: reddit (reddit:ClaudeCode)published 2026-11-01 20:55ingested 2026-11-02 02:41

retained text

Eight years as a DevOps engineer, running other companies' platforms, and too much YAML, feeling like rebuilding the same infrastructure with an over-complicated process. Three years ago I started building the one I wanted: give it a repo, get a URL. A Vercel-like in some sense, but driven by different technical approaches that try to optimize many aspects, and an opinionated simple UI to drive users towards best practices without them needing to dig into the technical side of DevOps.

The tooling changed a lot along the way. The beginning was by hand, laying down the architecture, optimizing code. Then ChatGPT quickly arrived and I was copying and pasting code, feeling like I moved twice as fast. Then Cursor, which edits the file for you. Looking back I feel like even with Cursor editing code and me reviewing each change carefully, I wouldn't have landed maybe a 10th of what I have today. Though it was already far faster than coding by hand (just reminding myself, while writing these lines, of the weeks of refactoring that would take 10 minutes nowadays).

And this year Claude Code. As a developer this is the moment where I got really conflicted about giving freedom to Claude: wanting to know how the code works and taking time to review, while seeing that I was the bottleneck and it could go even faster. Along the way I tried other fully vibe-coded projects to confidently discover how to use Claude, and shifted my approach towards testing every feature extensively, peer review with other agents, and hammering pre-production to extract as many errors as possible, doing end-to-end testing over and over. Nowadays I use this approach across many worktrees, while trying to keep at most 3 different features in parallel to limit the mental load.

People say AI builds a product in an afternoon, and for the demo that's true. The project still took months with Claude Code (thankfully I took pleasure in it), and the reason is that the happy path is maybe 5% of the work; the rest is that everything you implement adds more ways to fail (no different from classic coding, just shifted to another level of complexity). "One repo, one URL" sounds like one feature, but it crosses many systems (git, build, image registry, cluster, routing, database, disk, logs) and every feature I add touches them all. Each button hides a chain: "put it live" means build, health check, flip the route, retire the old copy, keep the data, and undo everything if a step fails. What is running lives in five places that drift apart, so you need checks, alerts, and a way to repair (most of my postmortems come from there). Unit tests are not enough; a deploy implies that each test is a real build on a real cluster, minutes each. Looking back, it's probably thousands of runs.

What it does today: every commit gets its own URL, its own running copy, and its own copy of the data, so you can go back to any version with its data. Agents can deploy straight from a webchat through MCP or through your local harness. Hundreds of self-checks compare what the platform says with what actually runs. It's live at [https://kad.dev](https://kad.dev) It's my project, made for solo devs and small teams who want their app online without doing the DevOps part. There is no paid plan right now; I'm just looking for users to improve the app with.

**Edit 1**:  I see people reacting more to the project (though I gladly welcome your criticism) than to the theme of the post, which questions the easy trend of saying we can build anything in 15 minutes. Building any real project still takes time.

**Edit 2:** Bringing up Vercel was a bad idea. It was just an easy way to understand the project, but if you want differentiating features, here are some:  
\- an opinionated infrastructure, so you can easily move data (SQL and S3) between commits (no more dev/preprod environments to maintain)  
\- you can manage everything though MCP. (It currently deploys itself though mcp)  
\- But of course you can do the same elsewhere with Claude, but run bash through SSH or GitHub is very forgiving: here there is automation and hard security wall to prevent it from destroying your production.

document fetched 2026-11-02 00:34 · canonical https://reddit.com/r/ClaudeCode/comments/1xle681/3_years_to_build_one_repo_one_url_claude_code