The best cheap mini PC for an AI agent,
this Memorial Day.

Most Memorial Day mini PC deals are written for someone who wants to play Helldivers on a TV. The same hardware is also the cheapest entry point to a 24/7 AI agent.

The pivot: same box, different ROI.

Kotaku's framing was the standard one. A Ryzen 7 mini PC at twenty percent off, contrasted against a Mac mini whose price is creeping the other way. Fine. Accurate. But it's the wrong question for an operator.

The question for an operator is: what's the cheapest box I can leave on for a year that runs an AI agent against my email, my CRM, my spreadsheets, my voicemail transcripts — and how much does the actual thinking cost me on top.

On a $400 box with an OpenRouter API key, the answer is "the box runs forever, and the thinking costs cents per task." A well-scoped business agent — triage inbox, summarize calls, flag follow-ups, draft a quote — runs in the low single dollars a day. That's a small fraction of what one missed call costs.

The hardware Kotaku wrote up.

The GMKtec M5 Ultra is a small fanned mini PC with an AMD Ryzen 7 7730U, 16GB of RAM (expandable to 64GB), a 512GB PCIe SSD (expandable to 4TB, with a second M.2 slot for more), and integrated AMD Radeon graphics. Wi-Fi, Bluetooth, the usual port salad, dual displays.

For an agent host, none of the gaming-relevant specs matter. The integrated GPU is irrelevant because we're not running inference locally. What matters is: it boots, it stays on, it has enough RAM to run a Python or Node process that talks to APIs and disks, and it doesn't draw 400 watts from your office plug.

16GB of RAM is plenty for that. The Ryzen 7 7730U is overkill for it. The 512GB SSD is overkill twice — agent logs, vector indexes, transcripts, all of it fits in well under 100GB for most one-shop use cases. You're not buying compute. You're buying always-on.

Why OpenRouter, not local inference.

We wrote a long field note in May about running a local model on a mini PC. That post is about a $2,000 box with 64GB of unified memory that runs a 70B-parameter model on-prem, no API calls leaving the building. It's the right call when privacy is the constraint or when the workload is heavy enough that token costs would compound.

This is the other path. A $400 box doesn't have the memory or the GPU to run a real local model — and most operators don't need to. They need an agent that thinks well, calls the right tools, and stays on. The cheapest way to get "thinks well" is OpenRouter: one API key, one library call, and you're routing the same agent code to Claude Sonnet, GPT-5, Llama, Gemini, DeepSeek — whichever model is best for the job that minute.

You don't pick a model up front and live with it. You pick it per task. Drafting a customer reply? A small fast model. Reasoning over a spreadsheet? A bigger one. Cheap recurring triage? An even smaller one. OpenRouter handles the routing, the failover, the billing in one place. It's the model layer an agent should have had from day one.

The stack, in plain English.

• The box. The M5 Ultra, plugged in, on a small UPS, with a static IP on your office network. Linux is fine. Windows is fine. Pick what you'll actually maintain.
• The runtime. Python or Node, your call. A simple process manager — systemd on Linux, pm2 on Node, nssm on Windows — so the agent comes back up if the box reboots.
• The model layer. One OpenRouter account. One API key in an env file. Every model call goes through one HTTP endpoint.
• The agent code. A scheduler that wakes up every N minutes, pulls the inboxes and queues it owns, decides what to do, calls tools, writes back. ~300 lines for a first useful one.
• The tools. Whatever the shop actually uses. Gmail API, a CRM webhook, a Twilio number, a Google Sheet. The agent is just orchestration on top of these.
• The log. Every run writes a row. You read it Monday morning and the agent gets one round of edits a week. That's the whole feedback loop.

No GPU drivers. No model weights to download. No fine-tuning. No vector database unless you actually need one. The agent on the box is plumbing. OpenRouter is the brain rental.

The cost math, at a year scale.

• Hardware. $400 one-time. Amortize over 3 years, $11/month.
• Power. ~20W average. About $25/year on a typical US plug.
• OpenRouter tokens. Highly task-dependent. A scoped business agent doing inbox triage, lead summaries, and quote drafts for one shop typically lands in the $30–$100/month range. Cap your spend in the dashboard.
• Software. Zero, if you write the agent yourself. Whatever you'd pay an integrator otherwise.

Call it $50–$120 a month, all in. The same business outcome from a cloud "AI assistant" SaaS — if one even exists for your workflow — is usually $200–$600/month with a per-seat model and a vendor lock you don't want. The mini-PC-plus-OpenRouter path is cheaper, more flexible, and the agent code is yours.

Honest tradeoffs.

This is not the rig if your data can't leave the building. Every model call still goes out to OpenRouter and on to a hyperscaler. If you're under HIPAA, under a DPA that names specific subprocessors, or just allergic to cloud, go read the local-LLM field note instead. Bigger box, no API calls, real privacy story.

It's also not the rig if you need millisecond latency on every call. OpenRouter adds a hop. For most business agents — the ones that run on a 5-minute cadence and act on email, voicemail, and CRM rows — you will never notice. For real-time voice or for a customer-facing chat, you'll want a different shape.

And the deal itself will end. The hardware argument doesn't. At $500 list, the M5 Ultra is still the cheapest "leave it on forever" box we'd ship today. At $400, it's a give-away.

Takeaway.

The cheapest way to put an AI agent on for your shop, today, is to buy a $400 mini PC, plug it into the wall, sign up for an OpenRouter account, and write the 300 lines of code that actually do your work. The compute is rented. The agent is yours. The bill is small.

Gamers will buy the M5 Ultra for the frame rates. Operators should buy it for the agent.