# What it costs to run a company on open-source agents > The bill for a company on agents comes from model calls, machine time and review minutes, not the platform fee. Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers. URL: https://companyrunbyagents.com/cost ## A credit, in the pricing page's words `kortix.com/pricing` ``` 1 credit = $0.01. 1 credit buys about 5 minutes of agent computer. Credits pay for the agent computer: the cloud machine where your agent runs code, browses and edits files. It bills per second while it runs and uses no compute while it is stopped. Each call costs the provider's base price plus a 5.5% platform fee, paid from credits. ``` Source: https://kortix.com/pricing Five lines that set the machine rate and the model markup you carry into your own row. ## The fee is the smallest number on the invoice A reader comparing agent tools puts the subscription next to the free option and picks the free one. That comparison leaves out the two lines that grow while a platform fee stays flat: the model calls each session makes, and the minutes a person spends reading what came back. OpenClaw's own page is direct about the first half. It describes itself as open source, running on your machine, with "No subscription. No hosted tier. No token." A company can install it for nothing and point it at a model provider. The invoice that follows belongs to the provider, plus whatever the machine costs. The published [seven-agent experiment](https://hellonunc.com/blog/posts/company-run-by-ai-agents/) shows the shape of that bill: seven agents on one 4GB Hetzner VPS, with the owner's summary of the infrastructure in one line, "A cheap VPS + LLM API calls". What it leaves out decides whether the number stays small: how many sessions run a day, and how long a person spends on each result. ## Three shapes, three bills Shape one is Kortix Cloud. Free is $0 with 500 credits per month, every feature and 1 project, on your own provider keys or a 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. Members are unlimited and free, and the account shares one credit pool. Shape two is the same product on your own metal. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. The servers, their power, and the people who patch them are yours, and so is every model call. Shape three is one agent on a machine you already own, which is the shape OpenClaw's page describes: no subscription, no hosted tier, no token. The software costs nothing. The VPS, the systemd units that keep it alive, the keys and the check-ins are yours. Two of the three charge no platform fee at all, and the third has a free tier. On that evidence the software looks nearly free in every direction. The difference moves to the machines and to the person reading the output, and neither of those appears in a pricing table. ## What a credit pays for Kortix prices in credits, and the page defines one: 1 credit = $0.01, and 1 credit buys about 5 minutes of agent computer. Credits pay for the agent computer, the cloud machine where an agent runs code, browses and edits files. It bills per second while it runs, and nothing while it is stopped. The default 2 vCPU / 4 GB / 20 GB machine comes to about 11 credits an hour. That meter follows sessions rather than people. One isolated sandbox per session. Each session has its own isolated machine and branch. Ten sessions that each run an hour bill ten machine hours; ten people sharing one session bill one. A stopped machine uses no compute, and its stored files are still billed as storage. A company that runs on triggers and short jobs pays for the work rather than for the wait. Credits pay for Kortix models as well, which is the route for a company that would rather not manage a provider key. That is where the platform fee actually lands, on the calls, and it is the smallest part of what a working session costs. ## Model spend per session, not per member Three routes, three bills. A Kortix model needs no key: each call costs the provider's base price plus a 5.5% platform fee, paid from credits, and the app shows the base prices. Your own API key sends the call straight to your provider, and that provider bills it. A connected ChatGPT subscription uses no credits at all, so per-call spending becomes a subscription you already hold, with the usage limits that come with it. What a session consumes inside any of those routes depends on what it reads. An agent that opens a repository, reads six files and retries a failed command spends several times what a one-line classification costs. Work that fans out across several agents multiplies the figure, since each agent carries its own context and pays for it. A long memory file costs money every time it is loaded. The number to track is the cost of a finished piece of work, a merged change request or a sent reply, rather than the cost of a message. Divide a month of model spend by the pieces that shipped and you have a figure you can put beside the alternative, which is a person doing the same work. ## Review minutes are the line nobody budgets The seven-agent experiment was reviewed once a day and still left half the agents' own files unupdated and two of the seven agents unnamed inside a week, as the [write-up](https://hellonunc.com/blog/posts/company-run-by-ai-agents/) records. A chat check-in has nothing to read, and a fluent reply says nothing about whether a memory file kept its last paragraph. Review does have a unit. When an agent changes the project itself, for example a skill or its memory, it opens a change request. A person reviews it on the Review page and merges it. The reviewer reads the lines that changed against the version that ran yesterday, so ten changes make ten reads, and the tenth read costs what the first one did. That is the line to put beside the model bill: changes a week, minutes a read, and the hourly cost of whoever signs off. The rate is usually the highest on the payroll, because a person allowed to approve a change to the company's own rules is rarely the most junior one in the room. ## Put your own numbers in a row ``` monthly = sessions x model spend per session + machine hours x what a credit costs + changes x review minutes / 60 x loaded hourly rate ``` The second line has a published rate, and the first has a published markup. Only the third comes entirely out of your own calendar, and no pricing page can quote it. Run all three over a month of real work, then compare the total with what the same month cost before the agents arrived. ## What each cost shape charges for Three ways to run the same agents, and the line each one leaves with you. | | Kortix Cloud (recommended) | Kortix self-hosted | OpenClaw | | --- | --- | --- | --- | | Platform fee | Free at $0; credits from $50 a month. | None. Self-host is free. | No subscription and no hosted tier. | | What you pay for | The agent computer and Kortix model calls. | Your own servers and their upkeep. | A VPS and model API calls. | | Self-hosting | Kortix Cloud, your VPC, or on-prem. | Free, on your own network. | The only option it offers. | | Who runs the machines | Kortix. | Your team. | You. | | Where model cost sits | In credits for Kortix models, or your own key. | Your provider key or subscription. | Your provider key. | | Review before work lands | A change request a person merges. | A change request a person merges. | The owner approves the big calls. | | Open source | Yes | Yes | Yes | Sources: the Kortix [pricing page](https://kortix.com/pricing), OpenClaw's [own page](https://openclaw.ai) and its [seven-agent write-up](https://hellonunc.com/blog/posts/company-run-by-ai-agents/), all read 4 November 2026. ## The four cost questions buyers ask Short answers, before you fill in the row. ### How much does it cost to run agents? Three numbers decide it: model spend per session, machine hours and review minutes per change. The platform fee is the smallest of them. Kortix Cloud starts at $0 with 500 credits a month, and self-hosted Kortix is free. ### Is self-hosting really free? Yes. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. You still pay for the servers it runs on and for every model call the agents make. ### What do Kortix credits pay for? The agent computer and Kortix models. The machine bills per second and nothing while it is stopped, and a Kortix model call costs the provider's base price plus a 5.5% platform fee. Your own key sends the calls to your provider. ### Do I need my own API keys? No. A connected ChatGPT subscription carries the model calls without spending credits, and the Free plan runs on your own provider keys or that subscription. A key from a model provider is another route, and that provider bills it. ## Start on the free tier and count your own month One project, 500 credits and a week of real sessions will tell you more than any pricing page. [Try Kortix](https://kortix.com) ยท [See pricing](https://kortix.com/pricing)