What self-hosted open-source AI agents cost to run

The platform line goes to zero

Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. A self-hosted instance runs with billing off by default (KORTIX_BILLING_INTERNAL_ENABLED=false), and it has no credits, no plan limits and no managed models. Nothing on that server sends Kortix a monthly invoice.

It helps to know what the managed account charges, because that is the number the decision usually turns on. Free is $0 with 500 credits per month, every feature and 1 project, on the user's own provider keys or ChatGPT subscription. The paid plan is credits: 25 tiers from $50 a month (5,000 credits) to $50,000 a month (5,000,000 credits), at 100 credits per dollar. Managed models are on the Credits plan only. Nothing on a self-hosted instance bills that way, which makes the managed line the one self-hosting deletes.

The server is the first line you add

A self-hosted instance is one Docker Compose stack on a Linux server with a public IP and ports 80 and 443 open. The published container limits are kortix-api 2048m, llm-gateway 2048m, frontend 512m and Postgres at a fixed 1280m. Add them and you get 5,888MB before the host itself needs anything. The preflight in the operator kit fails below 6,000MB, warns up to 7,500MB and passes above, so a 4GB machine stops there, before init writes a file.

Disk has a floor of its own. kortix-updater stops below KORTIX_MIN_FREE_DISK_MB, 2,048MB by default, and the instance keeps one previous release for rollback. A 6GB server clears the failure and sits in the warning band; 8GB clears both. Your provider sets the price of that machine, and no number here changes it.

Keep the split in view while you size it. The server holds the control plane, Postgres and the stored files. Sessions do not run there. kortix-api creates each one on a Platinum sandbox with your PLATINUM_API_KEY, and the sandbox calls back to the instance's public URL, so the machine you rent is sized for the stack and its database. The checks page shows what each record buys you.

The model line is your provider's invoice

A self-hosted instance has no managed models, so model calls go to the providers your projects connect. Any model provider with your own keys. Or the ChatGPT plan you already pay for. You add the key in the web app after the first sign-in, and each project keeps its own, which is why this line is however many provider accounts you choose to hold.

With billing off, the model line arrives as the provider's own invoice and nothing else. The managed plans are on Kortix.

The hours line is the one nobody quotes

Four jobs recur, and each one has a command in the kit.

• ./bin/backup.sh before every update and every env change. It stops the stack, archives the instance directory, and writes a sha256 plus a key fingerprint. .env holds API_KEY_SECRET, which decrypts every stored secret, so an archive without it restores a stack that serves pages and cannot read its own credentials.
• ./bin/doctor.sh before anyone signs in and after each update. Seven checks, and six of seven is the usual first result.
• The daily update window, 02:00 in America/New_York unless you change it. A failed migration swaps nothing and the old version keeps answering, so the cost is usually a morning rather than an outage.
• A restore drill on a spare instance. restore.sh refuses a key that does not match, and that refusal is better met on a calm day than during a recovery.

verify-all.sh in cron reruns sign-up, doctor, update status and backup age in one pass, and exits 1 when a person is needed. That is the schedule; the minutes are yours. A worked example with stated inputs: 15 minutes a week for the pass is 13 hours a year, and at $60 an hour the line is $780 before any restore drill. Put your own minutes and rate in and the number moves.

What the total is not

The subscription is the number everyone compares, and it is the one self-hosting removes. The three that stay are a server sized to a container floor, model calls priced by a provider, and recurring hours of someone's attention. Where model spend is small and nobody's hour is charged to anything, self-hosting lands close to free. Where a team's hours are real money, the monthly plan can be the cheaper invoice even though it is the only one with a price printed on it.

Before you size the machine, run the preflight. What the stack installs follows the same commands in order, and Read the docs for every variable the instance takes.