Honestly, if it weren't for the frontier models, I wouldn't be making tools and building things that really entertain me. I started out inside Gemini’s canvas back with the release of Gemini 3.0 (which was originally a 2.x version under a different name). While the current environment is full of supercar makers, if you want a local dev server with HTTPS for Progressive Web Apps, the standard answer is Vite paired with vite-plugin-mkcert or some React/Preact setup. If you want to orchestrate multiple AI agents in your codebase, you reach for deeply integrated tools like Cursor or Aider, or for multi-model workspaces, you lean on VS Code (or VSCodium) and environments like Antigravity. I genuinely love the IDE dev space, but it's like when Chrome used to be lightweight and extremely functional... now it's absorbing half of your system resources just to load a basic window.
I just don't like excess complexity that adds proprietary access without real value.
The Embodiment Problem
We don't suffer from the embodiment problem that AI models do. We can physically see, feel, and interact with what we build. We know when movement feels slightly sluggish, when a layout feels cramped, or when something is instinctively "off" just by looking at it.
I've had instances of extensive back and forth with models where I'd have to build a raw example (not just yap) to start getting things moving in the right direction. Language models can't experience the visual or tactile reality of the product. In fact, if you yap enough to these models, you'll notice they all eventually call you "the anchor." They do that to reinforce positivity: you're "tethering them to reality" and so on. I mean, isn't that just softening the blow of calling you out for being a drag and stopping the ships? Pun intended.
Whatever the case, hopefully this workspace setup can reduce that drag and bring reasonable semi-automation to your development.
The Deliberate Kit Car
Rather than adopting a bloated supercar, we built a highly deliberate, lightweight hybrid workspace from the ground up. It achieves the exact same results with absolute transparency.
Instead of installing an entire Vite ecosystem, we wrote a custom Node launcher. It handles
nmkcert CA installation, local HTTPS generation, dynamic port mapping, and
single-source Content Security Policies. It does all of this in under 200 lines of standard ESM
Node with zero external dependencies. It is completely self-healing. It just works.
Clean Boundaries for AI Orchestration
AI tools love to generate scaffolding. Heavy IDEs like Cursor often automate this directly into your working tree. We wanted total visibility and control over the process.
Our solution relies on a strict physical divide. The production repository remains pristine. All
AI prototyping, testing, and debugging happens strictly in a parallel, untracked
scaffold/ directory. To orchestrate different Language Models, we use a simple
markdown ledger: scaffold/comms.md.
When one agent (like Claude) finishes reasoning through a structural change, it writes an asynchronous handoff log. When the next agent (like Gemini) arrives for rapid execution, it reads the ledger and picks up the thread. The transition is seamless. There is no proprietary magic. It is just clean, transparent communication.
Native Abstraction
This philosophy extends to our data layer. Rather than pulling in heavy storage wrappers or desktop frameworks like Electron, our storage bridge adapts natively. It binds to the File System Access API in Chrome, routes to the Origin Private File System (OPFS) in Safari, and falls back gracefully to IndexedDB elsewhere.
By rejecting the supercars, we engineered a workspace that feels uniquely responsive. It proves you do not need excess complexity to build state-of-the-art applications. A highly disciplined, zero-dependency setup will often outperform a heavy framework, leaving the spotlight exactly where it belongs: on the work itself.
The industry is full of supercars. If you want a local dev server with HTTPS for Progressive Web
Apps, the standard answer is Vite paired with vite-plugin-mkcert. If you want to
orchestrate multiple AI agents in your codebase, you reach for deeply integrated tools like
Cursor or Aider.
These tools are undeniably powerful. They are also incredibly heavy. They pull in massive dependency chains. They obscure their core orchestration behind proprietary access models. They create a black box around your development process.