Why Do AI Agents Need a Server? Because They're Daemons, Not Apps

Axel Grubba, October 29, 2026
Let AI run your small business with Crevio
Crevio E-Commerce Platforms logo
Crevio
Sponsored
5.0
(1)
Free plan available
Crevio is AI that runs your small business: it builds your website, finds your customers, takes paymen... Learn more about Crevio
Get an AI summary of this post on:

An agent is a daemon, not an app.

That’s the whole answer, and once it lands the purchase decision makes itself. An application runs while you’re using it. A daemon runs whether you’re there or not, waits for things to happen, and acts when they do.

You can install an agent on your laptop today and it will work beautifully — right up until you close the lid.

What a laptop structurally can’t do

It sleeps. This is the one that matters, and it’s not a settings problem. A sleeping machine doesn’t fire scheduled jobs, doesn’t hold a messaging connection open, and doesn’t answer anything. Your agent’s 8am summary doesn’t arrive at 8am; it arrives whenever you next open the laptop, if at all. Every other item on this list is negotiable. This one isn’t.

It moves between networks. Home Wi-Fi, a café, tethering. Each transition drops long-lived connections, and while a well-written agent reconnects, “well-written agent reconnects” is a different property from “connection was never interrupted.”

It’s the machine you also use. An agent with shell access and your credentials, running on the laptop holding your email, your SSH keys and your browser sessions, has no isolation at all. On a server, a compromise gets an attacker a server. On your laptop, it gets them your laptop.

Its egress address changes. Every network change gives you a different public IP. If any API you call uses an allowlist, you’re re-adding addresses forever.

The inbound question, honestly

Most articles on this topic put “you need a stable inbound address for webhooks” near the top. For the common agent setup, that’s wrong, and it’s worth correcting because it makes people think the problem is harder than it is.

The two most-used channels connect outbound:

  • Telegram — its integration documentation states that “long polling is the default transport; webhook mode is optional.” Your agent asks Telegram for updates. Telegram never initiates a connection to you.
  • WhatsApp — connects as a companion device over WhatsApp Web, an outbound session your agent opens and holds.

So a default agent needs no inbound ports, no port forwarding and no static IP. It works behind a home router, behind CGNAT, on a machine with nothing exposed at all.

Inbound only becomes necessary when something outside genuinely has to reach you: a Stripe or GitHub webhook, a Slack event subscription, or Telegram if you deliberately switch it to webhook mode. That’s a real scenario — it’s just not the default one, and it isn’t the reason your laptop won’t do.

The reason your laptop won’t do is that it sleeps.

Where the half-measures break

Two workarounds come up constantly. Both are legitimate, and both solve a different problem than the one you have.

Tunnels — ngrok, Cloudflare Tunnel. These give a machine without a public address a reachable URL, by having it dial out to a relay. Genuinely useful when you do need inbound.

Where they break: a tunnel doesn’t keep your laptop awake. It solves addressing, not availability. Point a webhook at a tunnel to a sleeping machine and the requests arrive nowhere. Free tiers also hand you a URL that changes between sessions, which means re-registering it everywhere each time.

Leaving the laptop open and awake. It works. It also means a machine you want to close is running permanently, drawing power, spinning fans, and going offline every time you travel. People try this for about a fortnight.

Home hardware — a Raspberry Pi or a Mac mini — genuinely solves the awake problem, and is a reasonable answer. It just isn’t cheaper than a small VPS any more: Raspberry Pi raised prices twice in a year over memory costs, putting a Pi 5 8GB at $175. We did the 24-month arithmetic, and it lands close to an equivalent VPS while losing on uptime and isolation.

What “always-on” actually costs

Less than people expect, because the model isn’t running on your hardware — the agent orchestrates and calls a provider’s API, so you’re paying for a small always-on box rather than anything specialised.

What you’re running RAM Typical monthly
Text-only agent, hosted model API 4GB $5–10
Agent with browser automation 8GB $10–15
Agent plus a self-hosted model 8GB+ and a GPU Considerably more

The jump from 4GB to 8GB is browser automation, and it’s not marginal: a single headless Chromium instance wants 1–2GB entirely to itself. Undersize it and the failure is silent — the kernel kills the agent and nothing appears in its own log.

And you don’t need a GPU. That’s the most expensive misconception in this category, and it’s wrong for the same reason: the model runs on the provider’s hardware.

Which server

Hostinger is the simplest starting point — 8GB tiers and one-click agent templates, so you’re not building the machine before you can use it. Hetzner is the cheapest per gigabyte when its CX tier is in stock — worth checking before you commit, because that tier was showing as unavailable when we last looked, and the CPX line you’d fall back to is several times dearer. A plain Ubuntu box is all this requires. Cloudzy bills monthly with no term, which suits finding out whether you’ll keep it. Railway removes the server entirely — though its metered model bills for residency, and an agent holding channel connections is resident by definition, which we priced in Railway vs a VPS. UltaHost sits at the budget end; confirm the plan specs in writing.

DigitalOcean is a common choice here too, but its tracked link isn’t working on our end, so we’ve named it without linking.

Once you’ve bought one, the first ten minutes covers the setup order that stops you locking yourself out, and self-hosting an agent on a VPS does the install with hardening first.

How we checked this

That Telegram uses long polling by default with webhook mode optional, and that WhatsApp connects as an outbound companion-device session, are from the agent software’s own channel documentation, read in August 2026. Raspberry Pi pricing and the company’s stated reason for raising it are from Raspberry Pi’s own announcements. RAM guidance is the published floor used across our AI hosting coverage.

A correction worth stating. The brief for this article listed “a stable inbound address for webhooks from Telegram, Slack or Stripe” among the things a laptop can’t provide. For Telegram that’s not accurate on default settings — it polls outward — and the same applies to WhatsApp. Inbound addressing matters for genuine third-party webhooks like Stripe, but presenting it as a headline requirement overstates the difficulty of self-hosting and points people at tunnels when their actual problem is sleep.

What we did not do: we haven’t measured reconnection behaviour across network transitions or laptop sleep cycles. The claims above follow from what sleep does to a process and what the transports are documented to do, not from a test we ran.

The hosting links above are affiliate links.

FAQ

Can I run an AI agent on my laptop?

Yes, and it will work while the laptop is awake and connected. The problem is that a closed laptop fires no scheduled jobs and holds no connections, so anything time-based simply doesn’t happen.

Do I need a static IP for an AI agent?

Usually not. Telegram polls outward by default and WhatsApp connects as an outbound session, so the agent initiates every connection. You need inbound addressing only for genuine third-party webhooks such as Stripe or GitHub.

Will ngrok or Cloudflare Tunnel solve this?

Only the addressing half. A tunnel gives an unreachable machine a public URL, but it doesn’t keep that machine awake — webhooks pointed at a tunnel to a sleeping laptop arrive nowhere.

How much RAM does an always-on agent need?

4GB for a text-only agent calling a hosted model, 8GB once browser automation is involved, because a headless Chromium instance wants 1–2GB by itself.

Do I need a GPU?

No, unless you also run a model locally. The agent orchestrates on your server and the model runs on the provider’s hardware, which is why a small VPS is sufficient.

Is a Raspberry Pi a good alternative?

It solves the always-on problem, but it’s no longer clearly cheaper — a Pi 5 8GB is now $175 after two price rises, which works out close to an equivalent VPS over two years while losing on uptime and isolation.

Why not just leave the laptop open?

Because it works until you travel, close the lid, or reboot for an update — and because an agent with shell access and your credentials shouldn’t be sharing a machine with your email and SSH keys.

Founder & Software Review Editor
Axel Grubba is the founder of Findstack, a B2B software comparison platform, with his background spanning management consulting and venture capital where he invested in software. Recently, Axel has developed a passion for coding and enjoys traveling when he is not building and improving Findstack.
Business Software Reviews SaaS Product Evaluation CRM Software
Subscribe, get software deals straight to your inbox.
Join 9,300+ other entrepreneurs staying up-to-date on all the latest deals.
Zero spam. Unsubscribe at any time.