My AI stack: what I actually pay for

In short: my paid stack has one center of gravity, Claude, and a ring of tools around it that each earn their place. Claude Code runs my production day across multiple iTerm2 tabs, ChatGPT and Nano Banana cover different jobs, Fireflies sits in my meetings, and the most valuable tools on the list are ones I stopped buying and built instead.

The question behind the question

The question I get most often isn’t “should I use AI.” It’s “what should I actually pay for.” Fair question, because the tools multiply faster than anyone can evaluate them, and subscriptions have a way of accumulating quietly until you’re paying for a stack you don’t remember assembling.

So here’s mine, as it actually stands this month. Not a roundup of what’s interesting, just what’s on the card, and why it stays there.

The center: Claude, three ways

Most of my working day runs on a Claude plan, and it earns that by showing up in three different shapes. Claude Code in the terminal is the production floor: I run it in multiple iTerm2 tabs, usually one per project, and it handles the CMS builds, code embeds, audits, and content passes I described in how I actually build with AI. Claude Desktop is where the thinking happens before anything gets built. And Claude Cowork picks up the longer operational work around the projects, the multi-step tasks that aren’t code but still eat a morning. That’s my personal setup; the team at Dthree runs its own mix on client work.

Right now the heavier work runs on Claude Fable 5; it’s included in plan limits during Anthropic’s promotional window and moves to usage credits after, which I covered in the Fable 5 post.

The rest of the paid list

ChatGPT stays on the card as the second opinion. Different models have different tempers, and sometimes the fastest review of one model’s work is another model’s read of it.

Nano Banana, Google’s image model, covers the visual jobs that used to mean an hour in stock libraries: hero concepts, placeholder art, quick composites to test a direction.

Fireflies sits in my calls and takes live notes. That sounds like a small line item until you notice you’ve stopped losing decisions that were made out loud in meetings.

Ballpark, since people always ask: this personal AI layer, the Claude plan, ChatGPT, the image and meeting tools, lands somewhere around $80 to $150 a month depending on tiers, call it ₱5,000 to ₱9,000. For what it carries, which is most of my production week, it’s the cheapest line in the business.

The build layer, per project

Webflow and Figma are constants: design decisions still start in Figma, and Webflow is where most of the work my team and I ship at Dthree. Around them, Netlify, Vercel, and Supabase come in as the project calls for them, static demos on Netlify, app work on Vercel, a database when one’s needed. None of these are AI tools, but they’re where the AI’s output becomes something a client can use, so they belong on the bill. These scale with the work, so they live inside project budgets rather than the monthly card.

The tools I built instead of buying

Here’s the part of the stack nobody sells you. The CRM we run the agency on, the finance tool that tracks billing and payments, the project trackers: I didn’t subscribe to any of them. I built them with Claude, the whole team works in them every day, and they run on Claude intelligence through the API, summarizing threads and keeping the pipeline honest. The line item is API credits instead of a SaaS subscription, and the tool fits the way we actually work instead of the way a product manager guessed we might.

A year ago that would have sounded like a side project. Now it’s the most reliable part of the stack, because it changes when the business changes.

What stays free, on purpose

Google Analytics, Search Console, and Microsoft Clarity cover measurement without costing anything, and they earn their place precisely because they ask so little. The rule I use for free tiers anywhere in the stack: they’re fine until the day one costs a client an hour. That day, it graduates to a paid plan or leaves.

The test

I’ve bought my share of tools that felt like progress for exactly one weekend. The pattern was always the same: the tool was impressive, and nothing about my delivered work changed. So now every subscription has to answer one question before it gets added: does it remove a real step from work someone pays me for?

• If it makes production faster on billable work, it stays.
• If it makes me feel current but nothing ships differently, it goes.
• If I can’t tell after a month, that’s an answer too.

A stack isn’t a collection of good tools. It’s the shortest set of paid decisions between you and shipped work.

What this means for your stack

If you’re assembling your own, start from the work backwards, not from the tool list forwards. That’s the same argument I made in the stack you choose is the team you hire: every tool is a commitment to a way of working, and the monthly fee is the smallest part of the cost. Pay for what sits in the critical path, keep the experiments on free tiers, and when a category doesn’t fit anything off the shelf, consider that building it is now a real option.

And if that last part is the one you’re curious about, custom tools with real intelligence inside, that’s a conversation my team and I have with clients often, and I’m glad to have it with you.