Founderizer

Self-hosted AI agents: what has to be true

“Self-hosted” is said of setups that keep very different things on your side. Six questions show which parts of the work stay yours, and when a hosted platform is the better choice.

An AI agent is a model with tools, access and work to do. Self-hosting one can mean four different things: the agent runs on a machine you own, it uses a model account that is yours, the logins it acts with never leave your machine, and the record of what it did is yours to read. A setup can have one of these without the others, so the word alone tells you little. These are the questions that tell you the rest.

Where does the agent run?

On a machine you own, or on somebody else’s? A self-hosted agent runs on yours: a small Linux server, or a Mac that stays on. Not a laptop, which is shut at the hours an always-on agent is meant to be working. A €5 VPS is enough to begin with, because an agent spends most of its time waiting on a model and on the services it reaches, not computing. Which server should run your agents has the detail.

Whose model account does it use?

Running the agent on your machine does not put the model there. Most self-hosted agents still call a model provider, so ask whose account they call it with, and under whose terms. A fully local setup runs an open-weights model on your own hardware as well, which keeps the prompts on your side too, and needs a far larger machine than an agent that calls a provider.

Founderizer calls a provider. Each machine signs in to the Claude plan you already pay for, in your own browser, and we never hold a provider credential. The other side of that fact is that the plan’s limits are yours, and agents that run while you sleep can exhaust them. What always-on agents do to your AI plan says how to start without regretting it.

Where do the keys live?

An agent is only as useful as what it can reach: a mailbox, a payment service, a repository. On a hosted platform those logins have to be reachable from the platform’s machines, because that is where the work runs. A self-hosted agent can keep them in a vault on your own machine, and reach your code through a GitHub App you install in your own organisation, so that its vendor holds no token at all.

So ask any setup which of your logins its vendor could use if it wanted to. For Founderizer the answer is none: we hold an email address, a payment and an entitlement.

What is the record, and can you take it back?

An agent will sometimes be wrong in a way a person would not be. What matters then is whether you can see what it did and undo it. Here every grant and every owner decision is a commit in a git history on your machine, which you can read and revert, and anything an agent cannot do by itself reaches your phone as a question. Deciding what an agent may reach covers the grants themselves.

Who keeps it current?

This is the cost that self-hosting hides. Claude, Codex and Apple move every week, and a provider session on a machine lasts weeks, not years. Wire the agent yourself — a VPS, a Claude session, cron and some scripts — and you can have most of it working in a weekend. The week after is a standing obligation, and it is yours.

A maintained product moves that work to its vendor without moving your machine. Your hub takes a release when it is idle, checks itself afterwards, and returns to the version it pinned if a check fails. Wiring it yourself compares the two, including when your own scripts are the better choice.

What happens when you stop paying?

If the agents stop when the subscription does, the vendor is in their path, wherever the process runs. Ask whether there is a licence check, a call home or an expiry. Founderizer has none of the three: stop paying and your hub keeps running exactly as it is. What stops is ours — the updates, the app builds and the notifications to your phone.

When self-hosting is the wrong choice

When you do not want a server. It is a real cost and a real responsibility, and the backups, the disk and the bill are yours. When you want somebody else on call at three in the morning, or need to scale past what one small machine can do without thinking about the machine, a hosted platform is the better trade, and the comparison says so plainly. And when the work needs you in the loop anyway, Claude Code driven by hand is free with the plan you already have.

The checklist

  • It runs on a machine you own, and that machine stays on.
  • It signs in to a model account that is yours, and its vendor never holds that credential.
  • The logins it acts with stay in a vault on your machine.
  • Each agent gets only the access its job needs, one grant at a time.
  • What it did is a record you can read and revert.
  • An update checks itself, and goes back if a check fails.
  • Nothing switches it off when you stop paying.

This page is written by the company that sells Founderizer, which is built to answer yes to every line above. So hold Founderizer to the list as hard as anything else you consider, including a setup you wire yourself.

Questions

What is a self-hosted AI agent?

An AI agent whose process runs on a machine you own instead of on a vendor’s. Most still call a model provider for the model itself, so the questions that matter next are whose account that is, where the logins the agent uses are kept, and whether anything of the vendor’s sits in its path.

Do I need a GPU to self-host AI agents?

Not unless the model runs on your machine too. An agent that calls a provider spends most of its time waiting, and a €5 VPS is enough to begin with. Founderizer works that way, signed in to your own Claude plan.

Is a self-hosted agent safer than a hosted one?

It keeps your logins and your record on your side, which removes one party that could lose them. It does not remove the agent’s own mistakes: an agent acts with the access you gave it. Grant the access the job needs and no more, and start with the jobs whose mistakes are cheap.

The other guides

What it costs, what the agents actually do, and what you carry before you pay. Something wrong on this page? support@founderizer.com.