Why I’m using boring tools to build with AI

The more I use AI coding agents, the more convinced I am that the durable advantage is not a single tool, but a secure, model-agnostic development environment that can absorb whichever model is best this month.

Tech has been a hobby for decades. I built my first computer as a teenager and learned the quirks of computers and networking by scripting custom game levels and taking my bulky 90s tower PC to gaming parties with friends. For years, my limiting factor was my own coding ability. LLMs becoming good enough to write stable code has changed that. Like many others, I have become increasingly focused on cost-effective ways to build and maintain smaller niche apps — tools that solve real problems for me and my clients, but sit in markets too small to support full technology companies.

What my first AI coding experiences taught me

My first experiences with agentic coding were with Claude Code. It was impressive and built a working prototype remarkably quickly, but it burned through my subscription usage in no time at all. A couple of notifications from my Monzo business card in quick succession for usage top ups stopped me in my tracks! It was immediately apparent this would not be a financially sustainable way to build and maintain software (at least not for me anyway).
My next experience was with Replit. It made building remarkably easy and was considerably cheaper than Claude Code. Publishing on Replit’s servers would have been trivially easy, but I wanted to own the code, and control the hosting and the security environment, so I wanted it in my GitHub Enterprise account and hosted on Azure. That was awkward to configure and maintain, and it was always just a bit too painful to keep GitHub and Replit in sync; it felt unnecessarily complex.

The principles I settled on

These two experiences were informative, and helped me crystallise some principles that I wanted to follow if I was going to build and maintain software products:

Model agnosticism is necessary for an enduring setup. The frontier is changing constantly, and open-source models are improving quickly, as shown recently with GLM-5.2. I don’t want the app layer to restrict which models I can use. Building too tightly around Claude Code, Codex or any single harness feels strategically unwise, because their features are spreading to competitors almost as quickly as they are released. Over time, the usual cost and simplicity pressures of enterprise IT are likely to consolidate and rationalise the market.
Security must be robust. I needed a professional setup I can duplicate for clients that will withstand diligence scrutiny, not something hacked together.
It must be cost effective.
24/7 operation is necessary. The future will include LLM-powered autonomous agents that you can delegate to, such as Microsoft Scout, Microsoft’s enterprise-secure Autopilot agent built on OpenClaw technology, and they need compute 24/7.

The setup I now use: a secure, always-on dev box

I now have an Intune-managed, headless Mac mini sitting on my desk, available 24/7 and accessible from anywhere via Tailscale. It currently acts as a dev box, running GitHub Copilot in VS Code, with GitHub CLI installed for the Copilot agent and OpenRouter connected for model choice. The GitHub Copilot desktop app is there as a fallback, and I also have it on my MacBook Air.

Agentic development now feels straightforward. I can switch models from session to session, moving between frontier and open-source models via OpenRouter. Using GitHub CLI, the GitHub Copilot agent can keep local branches aligned with GitHub and open PRs for review. I also have private Azure infrastructure behind SSO, so I can test thoroughly on infrastructure that mirrors production without creating public exposure before deployment.

This is actually a pretty standard development setup used by engineers all over the world (although they often run it on their primary device), which I think is indicative of how fast things are changing; one minute everyone is experimenting with new tools; the next, the established tools have absorbed much of the new functionality.

Why this matters for business owners

For business owners, the value is not simply that small apps can now be built faster. It is that useful internal tools, workflow automations and niche software ideas can be explored without immediately committing to a full product team, a large vendor contract or a risky shadow-IT experiment. The right setup gives business owners a controlled way to use agentic development to prototype, test and operate software on familiar infrastructure, with clear ownership of the code, sensible security boundaries, and enough model flexibility to avoid being locked into whichever AI tool happens to be fashionable this month.

The unexpected benefit: fewer subscriptions

The final part of taking this model-agnostic approach is that I find myself drawn back towards Microsoft 365 Copilot. As Microsoft adds more model choice into Copilot experiences, including OpenAI and Anthropic models where supported, it becomes harder to justify maintaining several separate AI subscriptions for everyday work. I have already largely stopped using my separate ChatGPT and Claude subscriptions. Claude’s Design tool is still useful, but the designs it makes are just web pages that one prints to PDF. The same setup that I just described to make and maintain software is already proving capable of designing slides and marketing collateral. At this point I will be surprised if I renew my ChatGPT and Claude subscriptions at the end of their annual terms.

Bottom line

AI tools are moving incredibly quickly. The durable value is not in betting everything on one model, one app or one vendor wrapper. It is in owning the workflow, infrastructure and governance that let you take advantage of whichever tool is best next.

If you’re trying to make sure your team are well setup to take advantage of the benefits of AI, we can help. Reach out below.👇