August 11, 2026. xAI ships Grok Bot - a team of always-on AI agents that live in the cloud, keep working after you close your laptop, and can drive a browser, a filesystem, and a terminal on your behalf.

Genuinely impressive. Also a subscription.

Full access sits behind SuperGrok Heavy at $300 a month, or Cursor Ultra at $200, or Cursor Teams Premium at $120 a seat. Every agent you spin up gets, in xAI's own words, "its own persistent cloud computer, its own logins."

Translation: your credentials now live on a machine you don't own, rented at a price that resets every 30 days, forever.

So here's the question nobody asked during launch week, because launch weeks don't reward questions:

What are you actually paying $300 a month for?

Not the intelligence - Grok's intelligence is available a la carte through the API for a few dollars per million tokens. Not the "always-on" part, either. A $12-a-year electricity bill keeps a small computer running 24/7 without breaking a sweat.

What you're really renting is xAI's cloud infrastructure. And for most of what people actually want an "always-on agent" to do, you don't need it.

So I built the alternative. Grok as the mind, a Mac mini as the body.

Grok's live X search and frontier reasoning handle the 20% of tasks that genuinely need the internet in real time. A small local model on the Mac mini - tucked in a closet, drawing about as much power as a phone charger - handles the other 80%: monitoring, filtering, drafting, formatting, remembering. The cloud only gets a phone call when it's actually earning its keep.

This isn't a takedown of Grok Bot. It's a refusal to pay cloud-computer prices for tasks that never needed a cloud computer in the first place.

01 - What Grok Bot Actually Is

Strip away the launch-day excitement, and Grok Bot is a specific, well-engineered thing. Nothing vague about it:

Each bot runs on its own persistent cloud instance with browser, file system, and terminal access
You authenticate once through a secure handoff screen; the bot reuses those credentials going forward
A "teach a task" mode lets you demonstrate a workflow once, which the bot then repeats - including on a recurring schedule
It shipped simultaneously on macOS, iOS, Windows, and Linux, with full session takeover from your phone
Early testers used it for things like 30-minute social monitoring sweeps, Gmail triage, and Salesforce navigation via full browser control

hat's real, useful automation. No argument there.

The friction shows up in the pricing - and it showed up fast. Within days of launch, marketer Morgan Linton had to publicly walk back the confusion his own tweet caused: "I think ppl are confused about Grok Bot pricing... Quick reminder, it's $60/mo, not $300."

A builder correcting his own pricing tweet, in public, days after launch. That alone tells you the plans aren't simple. And $60 still doesn't buy the uncapped SuperGrok Heavy tier that most of the "always-on agent army" demos were actually running on.

Notice the bottom row. Nobody was talking about it the week Grok Bot launched - too busy being dazzled by the demos.

It's also the only row where you end up owning what you build.

02 - First Principles: What Does "Always-On" Really Require?

Break "always-on AI agent" into its actual engineering requirements and the magic evaporates fast. It's four boring pieces:

Something that stays powered on. Doesn't need to be a data center. A Mac mini idling as a home server pulls roughly 12 watts on average - about $1–2 a month in electricity in most of the US.
Something that can reason and act. This is the part that genuinely benefits from Grok - especially its live X Search tool, which does keyword, semantic, user, and thread-level search directly against real-time X data, filterable by handle (up to 20), date range, and media content, with citations attached. No other model has that lane into X natively.
Something that remembers. State, logs, drafts, embeddings - this is pure storage and retrieval. A local SQLite database or vector store on a Mac mini does this instantly and never leaks to a third party.
Something that runs the routine 80%. Formatting a reply, checking whether a mention matches a filter, summarizing a thread you already pulled - this is exactly what a small local model, run through Ollama or Apple's MLX framework, is built for. It doesn't need frontier intelligence. It needs to be fast, free per call, and private.

Only step 2 needs the cloud. That's it - one out of four.

Steps 1, 3, and 4 are solved problems, and a $599–$899 machine on your desk handles them better than any rented cloud computer ever will. No metering. No per-action billing. No vendor who can change the rules on you next quarter.

03 - The Machine: What the Mac Mini Actually Costs Now

Here's what makes the timing genuinely strange.

The Mac mini M4 launched in October 2024 at $599 for the base 16GB/256GB configuration. That price didn't survive 2026. Apple quietly pushed the same configuration to $899 earlier this year - and then, in August 2026, the exact same month Grok Bot launched, retired the M4 entirely and replaced it with the M6 and M5 Pro lineup.

Translation: the $599 machine this whole idea is built around is already gone from Apple's own store.

Gone, but not extinct. It's still out there - used, refurbished, open-box, on clearance - and it's findable for buyers who move now, right as demand for a cheap always-on AI box is about to spike because of Grok Bot's own launch.

That's the trend underneath the trend. Everyone's chasing the $300-a-month cloud agent. Almost nobody's noticed the $599 hardware tier just became a closing window.

04 - Local Memory Rule: Match the Box to the Model

Before you buy anything: know what your memory tier actually buys you locally. This is exactly the part most "just buy a Mac mini" threads skip:

Memory rule, simplified: 8GB ≈ a 7B model. 24GB ≈ a 32B model, comfortably. 128GB ≈ 70B territory.

For a Grok-hybrid bot, 24GB is the sweet spot - enough to run a capable 13B local model for routine work while leaving headroom for the background daemon, without ever touching swap.

05 - The Build: Grok Mind, Mac Mini Body

The architecture is deliberately boring, which is the point - boring things don't break at 3am.

The router script - the piece that decides "local or Grok" - is the entire product:

Keeping it alive 24/7 without a subscription is a solved macOS problem - launchd, not a cron job, not a dashboard:

Load it once -

Then pair it with two settings in System Settings → Energy: enable "Wake for network access" and disable sleep on power adapter. That's the entire "always-on" infrastructure Grok Bot charges $300 a month to abstract away.

06 - The Math: Receipts

Assume a moderately active bot: routine local calls running all day long, plus a realistic cloud workload of 8M input / 3M output tokens a month, split between Grok 4.1 Fast for quick lookups and Grok 4.6 for anything that needs real reasoning.

Round generously - a bit of headroom for setup time, an occasional heavier month - and you land around $89 in month one, under $40 a month after that.

Compare that to $300 a month for SuperGrok Heavy. Indefinitely. Forever. With no path to it ever getting cheaper.

07 - What You Actually Give Up

Fairness first: let's be honest about the trade.

You lose the polished mobile session-takeover UI. You lose the managed browser environment with saved logins across sites. You lose the "teach it once by demonstration" recorder - real engineering, the kind a weekend project won't replicate.

If your actual use case is "control a dozen SaaS logins from my phone while I'm at the gym," pay the $300 and don't look back. That's exactly what the product is for.

But that's not most people's use case. Most people want a bot that watches something, remembers everything, and only reaches into the cloud when it genuinely needs Grok's live X access or frontier reasoning.

For that, you don't need to rent the whole cloud computer. You need about 20% of it, on demand, at API rates.

08 - Real Posts Worth Pulling Into This

These are genuine, live posts on X from the last few weeks that back up every angle above - worth screenshotting straight from the source for the thread:

xAI's own announcement that the Live Search / x_search tool shipped free in beta for developers - the feature this entire hybrid depends on. @xai

Morgan Linton's public pricing correction - proof the $300 number confused even people building on top of it. @morganlinton

Austen Allred on why Grok Bot had to "just work" for consumers instead of requiring people to "install a server (or a Mac mini)" - the exact tension this article leans into, from the opposite side. @Austen

Peter Yang's side-by-side comparison of Grok Bot against other agent platforms, useful context for anyone deciding which way to go. @petergyang
@amuse's tongue-in-cheek take on selling off local Mac Minis now that Grok Bot exists - the counter-argument, worth including for balance and because it's genuinely funny. @amuse

09 - The Bottom Line

Every AI company wants the same thing right now: turn intelligence into a recurring subscription instead of a resource you pay for by the drop. Grok Bot is a good product built for that model. It's not wrong to want it.

But "always-on" was never the expensive part. A machine that stays on is one of the cheapest things you can buy. The expensive part is intelligence - and intelligence, right now, you can already buy by the token. No subscription required.

Own the box. Rent the brain only when you actually need it.

That's the whole idea. And right now - with the original $599 Mac mini already discontinued and a brand-new $300-a-month agent product barely three weeks old - is exactly the moment this trade is most obviously worth making.

If you read this far

Bookmark this.

You're not going to remember the router script, the launchd plist, or the exact pricing table when you're actually sitting in front of a Mac mini at 11pm trying to get this running. Save it now, use it then.

Follow for more.

This is one build. I'm going deeper - real benchmarks on the M6 vs. a used M4, whether MLX actually beats Ollama for this workload, and what breaks the math the day xAI ships Grok 5. Numbers over hype, every time.

Sources: pricing and specs current as of August 30, 2026 and subject to change - verify against Apple's store and xAI's API pricing page before publishing exact figures.

Grok Bot Costs $300 a Month. Mine Runs on a $599 Mac Mini for $89.

Job breakdowns