Using AI to hunt homelab deals in South Africa

South Africa has an import duty, a weak rand, and thin local stock, and the three of them together mean you pay two to three times what someone in the EU or US pays for the same box. A new small-form-factor PC with enough CPU and RAM to be a useful k3s node runs somewhere between R15,000 and R28,000 here. That’s not a homelab price. That’s a “maybe if I skip rent, the car payment, and insurance this month” price.

So the answer, like it usually is, was secondhand. Marketplaces like Takealot, Amazon.co.za, and Bobshop all carry refurbished and lease-return corporate hardware — the exact Dell OptiPlex Micro and Lenovo ThinkCentre Tiny boxes that make perfect k3s nodes once someone’s three-year lease is up. Finding a good deal, though, takes some digging: listings aren’t consistently named or categorized, and the actual bargains sit ten or more pages deep behind noise. Paging through that by hand is exactly the kind of tedious, mechanical work I didn’t want to do myself, so instead of grinding through it page by page, I handed the hunt for refurbished hardware to something that wouldn’t mind the grind.

Pointing Claude at the listings

There’s no pipeline here, no n8n workflow, no scheduled scraper. I used the Playwright MCP plugin and gave Claude one instruction: “I need a good SFF PC for my homelab.” That’s it. It went and browsed Bobshop itself, paged through the listings, and came back with candidates.

The one adjustment I made wasn’t a correction — the first pass wasn’t wrong, it just included sellers with rough scores mixed in with the good ones. I told it to only surface listings from better-rated sellers, and from a buyer’s perspective that immediately felt safer. Bobshop’s listings don’t tell you which sellers are worth trusting; an agent that can check ratings before surfacing a result does that filtering for you.

The haul

Three machines, all destined to be k3s nodes:

Dell OptiPlex 7060 Micro — i5-8500T, 8GB RAM, 256GB SSD, Windows 11 Pro, 1-year warranty — R3,249.97, marked 35% off the listing’s own reference price of R5,000.
Lenovo ThinkCentre M720q — i7-9700T, 8GB RAM, 256GB NVMe, Grade A Microsoft Certified Refurbished, 6-month warranty — R6,495, 28% off a R8,995 reference.
Lenovo ThinkCentre “M720q” — i7 8th-gen, 16GB RAM, 256GB NVMe, 6-month guarantee — R4,790. This one turned out to actually be an M920q when it arrived. Same seller mislabeled it in the listing; I got the newer, better board for tiny-PC money.

Plus three RAM sticks bought along the way to push each box past its stock 8GB — a 32GB Klevv DDR4 SO-DIMM, a 16GB SK hynix SO-DIMM, and a 16GB Samsung SO-DIMM — R1,200 to R1,800 each, because the whole point of buying used is spending the savings on the upgrade instead of the base unit.

Total spend across three machines and the RAM: a little under R19,000 — close to what one new equivalent machine alone would cost at the top of that R15k–R28k range.

The one hiccup, and why it didn’t matter

The 32GB Klevv RAM stick shipped to the wrong customer, and I received a DDR5 module that doesn’t even fit the boxes it was meant for. I contacted the seller the same day it arrived; they arranged a pickup, and the correct part was in my hands the next day.

That’s the actual payoff of filtering for seller rating instead of just price. Deals from unrated or poorly-rated sellers are the ones where a mix-up like that turns into a week of back-and-forth or a lost R1,800. A well-rated seller fixed it in about eighteen hours without me having to escalate anything.

The takeaway

You don’t need a scraper pipeline, a scoring model, or a scheduled job to make this work. Point an AI agent with a browser at a marketplace that’s too disorganized for you to want to page through yourself, tell it what you actually need, and let it do the tedious part. A homelab that would cost R45,000-plus in new hardware came in under R19,000 in refurbished corporate lease returns — a genuinely decent cluster, for a lot less than you’d assume it costs.