I Spent $20,000 on AI Tools, Here's What Actually Works

ELI5/TLDR

A creator’s tour of the AI stack he actually pays for. The headline number is theatre — the real list is short: Claude Max ($200/mo) for everything chat-and-agent, Conductor (free wrapper) to switch between Claude Code and Codex in one window, n8n self-hosted on Hostinger to move data around, Retune for visual tweaking before letting the agent commit, and ElevenLabs + Midjourney for voice and image fixes. Useful inventory if you’re building the same kind of one-person shop. Light on hard ROI numbers.

The Full Story

The big two, treated as commodities

Oliur’s main setup is Claude Max at $200/month, which he says he’s never hit the limit on despite leaning on Opus more than he probably should. The pitch isn’t about a particular model — it’s about the desktop app collapsing chat, “co-work” (file-system agent), and Claude Code into one window. He uses co-work for the things that used to mean opening five tabs: pulling fields from financial PDFs, comparing legal docs, organising a downloads folder, dragging data out of Linear or Google Drive into a CSV.

Codex gets equal billing. His rule is honest:

If there’s something I can’t do in Claude Code or Claude Code is struggling, then I switch to Codex. And then if Codex is struggling, I switch to Claude Code… I am not tied to any of these things. If a better model came out tomorrow, if a better app came out tomorrow, I would completely switch.

That’s the whole governing philosophy of the video. No religion about vendors.

Conductor — the wrapper that wraps the wrappers

The most actionable pick is Conductor: a free desktop app that hosts both Claude Code and Codex behind one UI, with projects on the left, chat in the middle, file diffs and terminal on the right. He calls it a graduation away from running Claude Code in the terminal — partly an aesthetic preference (he’s a designer), partly the ability to drag-drop images and visually verify what got attached. You still pay both subscriptions; the app itself is free.

The plumbing layer: Hostinger + n8n

Hostinger gets a slot not because it’s AI but because it’s the cheap, capped-bill VPS he runs everything on. The selling point he keeps returning to is the cap — flat $3.99/$7.99 plans, no surprise four-figure overage horror stories. On top of that he runs n8n self-hosted (Zapier/Make-style automation, but you own the box). His main use case is moving thousands of customer emails between systems every month — boring data shuffling that would otherwise eat his week. He flags one community template that’s a clean illustration: drop a receipt into Google Drive, n8n extracts vendor/date/total/tax with an AI step, writes it to a Google Sheet. Add WhatsApp on the front and you’ve got expense tracking that takes a photo.

The design and media side

Three smaller picks:

Retune.dev — a visual layer (think Figma) on top of your Claude Code or Codex project. You tweak padding, radius, font sizes live in the browser, then paste the diff back into the agent. The non-obvious benefit he names: it saves tokens, because you batch a hundred small UI changes into one prompt instead of round-tripping each one.
ElevenLabs — used for voice repairs in already-recorded videos. He’s patched in AI-generated lines in his own cloned voice and says no viewer has ever called it out. Limited to short fixes, not full narration.
Midjourney — the bragging-rights anchor. He generated an “acrylic paint” wallpaper pack entirely in Midjourney and says it’s earned tens of thousands of dollars on his store. The lesson isn’t about Midjourney specifically — it’s that image-gen is now good enough to be the product, not just the marketing for the product.

What’s missing

The “$20,000” framing is doing a lot of work. The actual recurring stack he describes adds up to roughly $200 (Claude) + maybe $20 (Codex) + $20 (ElevenLabs) + $30 (Midjourney) + ~$8 (Hostinger) = under $300/month. The 20K is presumably the historical churn of testing tools that didn’t make the cut — but he doesn’t list the rejects, which is the more useful video.

Key Takeaways

Claude Max ($200/mo) — chat + co-work (file agent) + Claude Code in one app; he’s never hit the Opus limit
Codex — kept as the sidecar; switch when Claude Code is stuck and vice versa
Conductor (free) — desktop wrapper that hosts both Claude Code and Codex with diff/terminal view; the recommendation if you’re tired of the terminal
Hostinger VPS ($4–8/mo) — capped-bill hosting, one-click installs for n8n and similar; chosen mostly to avoid runaway cloud bills
n8n (self-hosted on Hostinger) — Zapier/Make replacement; his use case is bulk email/customer data movement; receipt-to-sheet template is the killer demo
Retune.dev — visual UI editor that lives on top of agent projects; biggest win is batching changes to save tokens
ElevenLabs — voice cloning for short fixes inside already-recorded videos; full narration not yet there
Midjourney — sold a wallpaper pack made entirely with it for “tens of thousands”; image-gen as inventory, not marketing
The real philosophy — no brand loyalty, switch the moment something better ships

Claude’s Take

Clickbait alert: the “$20,000” is a thumbnail prop. The actual stack he’s running costs under $300/month. If you came for the shocking number you’ll leave annoyed.

What’s useful is narrower than the title suggests: the Conductor recommendation (genuinely a nice wrapper if you flip between Claude Code and Codex), the n8n + Hostinger pairing as a cheap automation backbone, and the Retune token-saving trick. The rest — Claude, Codex, ElevenLabs, Midjourney — is stuff Shantum already knows about and uses. There’s no comparison data, no failure stories, no “I tried X and it was rubbish, here’s why.” The video is more inventory than analysis.

Score: 5/10. Useful as a cross-check that you’re not missing something obvious. Not useful as a workflow blueprint. The Conductor + n8n pieces are the only two things worth following up on.