Skip to content
← All notes
5 min

What a self-hosted AI agent actually costs to run

Real monthly numbers for hardware, model API usage, and the costs people forget — for personal use and for a small team.

The software is free and open source. That is the part everyone hears, and it leads people to assume the whole thing is free. It is not — but it is cheaper than most of the SaaS subscriptions it replaces, and the numbers are knowable in advance.

Hardware: $0 to $799, once

The agent needs a machine that stays on.

  • A spare machine you already own — $0. An old laptop with the lid-close behaviour changed is genuinely fine. This is what we recommend to most people starting out.
  • A new Mac Mini — $499 to $799. Quiet, low power draw, and it will sit in a cupboard for years. The usual choice for people who decide they want this permanently.
  • A VPS — $5 to $20 a month. No hardware at all. The trade-off is that your agent’s data lives on someone else’s computer, which for some people defeats the purpose.
  • A Raspberry Pi 5 — around $80. Works, with 4GB RAM as the practical floor. Slower, and fiddlier to maintain.

Model API: the real recurring cost

This is where your money actually goes, and it varies by how much you use it.

  • Light personal use — a few dozen messages a day: $8 to $15 a month.
  • Regular personal use — 50 to 100 messages a day, some automated workflows: $15 to $25 a month.
  • Heavy personal use — always-on triage, scheduled research, long context: $25 to $50 a month.
  • Small team, 5 to 15 people$50 to $300 a month, depending heavily on what you automate.

Two things move this number more than anything else. Model choice: routing routine work to a cheaper model and reserving the expensive one for hard tasks can halve the bill. Context discipline: an agent that re-reads a huge history on every message costs several times one that does not. Both are configuration, and both are worth getting right at setup rather than discovering on your first invoice.

The costs people forget

  • Your time. For a technical person, expect 15 to 40 hours to get to a properly configured, hardened setup. Double it if you are not comfortable in a terminal. This is the single largest hidden cost and the main reason setup services exist.
  • Electricity. A Mac Mini running continuously is roughly $15 to $30 a year. A desktop with a discrete GPU is considerably more.
  • A domain, if you want webhooks — around $12 a year.
  • Patching. Not a cash cost, but a real one. This software ships security advisories regularly and someone has to read them.

A realistic first year

Reusing a spare machine, moderate personal use, doing the setup yourself:

Around $240 for the year, plus your own time.

Buying a Mac Mini and having it set up professionally at our founding rate:

Around $1,000 for the year, most of it the one-off hardware and setup.

For comparison, that is roughly what three or four mid-tier SaaS subscriptions cost — and unlike those, the machine and the data stay yours.

Where we make money, so you can judge the advice

We charge for setup, workflow builds, and maintenance. We take no margin on your API spend — that bill goes from you to the provider directly, and we configure caps so it cannot run away from you.

We mention this because cost advice is worth exactly as much as the incentives behind it. Ours is that a client whose bill surprises them is a client who cancels.


Want a real number for your situation rather than a range? Book a strategy session — working that out is most of what the hour is for.

Not sure which of these you need?

Start with a strategy session. An hour, $125, and you leave knowing what an agent would do for your week, what it would cost to run, and whether it is worth doing yet — including if the honest answer is not now.