Managed vs self-hosted AI agents: an honest comparison

Self-hosting an AI agent looks cheap. The VPS is a small monthly bill; the agent software is open source; you can be up and running in an afternoon. So why pay anyone to do it for you?

It's a fair question and we'll answer it honestly — including the cases where self-hosting genuinely is the right call.

The shape of the costs

What you see on the price page

VPS rental, model provider API spend, storage for backups, a domain. Add it up: $50–$200/month for a serious setup. That's the bill people anchor on.

What you don't see

The hidden costs are time and risk, not money:

• Initial setup that's actually correct.Not "it runs." Hardened SSH, firewalled, backed up, monitored, with the agent guardrails on. The first time you do this it's a long week. The fifth time it's a couple of hours. There is no fifth time for most people.
• Patches and updates. Both the OS and the agent itself move. Skipping patches is how servers get compromised; updating the agent without a snapshot+rollback path is how you wake up to a broken bot. Doing this safely takes a working update pipeline, not just apt upgrade.
• Backups, and the restore drills you have to run.See our other article. Untested backups are a hope, not a backup.
• Re-pairings and re-auths.OAuth tokens expire. Tailscale keys rotate. The messaging platforms occasionally invalidate sessions. Each one is a 20-minute fix — if you know the procedure cold and have the runbook handy. Otherwise it's an evening.
• Being on call.Agents talk to your customers. If one stops responding at 9pm on a Sunday, someone has to notice and fix it. "Someone" means you.

None of those are dollar costs. All of them are real.

When self-hosting is genuinely the right call

We'd tell you to self-host in any of these situations:

• You're a technical operator who enjoys the work.If running a hardened Linux box and writing the runbooks sounds like an interesting weekend, you don't need us. You'll end up with something close to what we'd build for you.
• The use case is personal or experimental.If the agent breaking for a day is a non-event, the operational overhead isn't worth it to anyone.
• You have an in-house SRE team with capacity.The marginal cost to your team is small if they're already doing this kind of work.

When self-hosting is a trap

• The agent is customer-facing.Downtime is customer downtime; security incidents are customer trust incidents. The cost of an outage isn't the VPS bill — it's the customers who saw it.
• You're a small team without a dedicated ops person.Operating this stuff competes with everything else you should be doing. Most weeks it's fine; the weeks it isn't are the ones it always isn't.
• You want to focus on the use case, not the plumbing.The interesting part of an AI agent is what it does for your business, not how it gets patched on Sunday night.

What managed gets you

Concretely:

• The same kind of single-tenant box you'd set up yourself — but built from a known-good template that's been through audit and refinement.
• A managed update path with snapshot-first safety. If the agent's new version regresses, we roll back. You don't notice.
• Nightly encrypted backups with monthly restore drills on a disposable box. The procedure isn't theoretical.
• Monitoring that pages someone other than you when the gateway stops responding.
• A single number to call when something needs fixing.

The pricing pitch in one sentence

If your time is worth more than the additional cost of a managed deployment, and the agent matters to your business — let someone else do the plumbing. If it isn't, or the agent doesn't, the open-source path is genuinely fine. We'd rather you choose clearly than convert at the wrong moment.