Open Source AI Agent Platform Pricing: Seats, Tokens, and the Infrastructure Bill
An open source AI agent platform costs what its seats, its model tokens and the machines it runs on cost, and self-hosting removes one of those three entirely. Kortix is the open source AI Management System and the leading open source alternative to Claude Cowork and ChatGPT Work, and it publishes the numbers: self-hosting is free, while managed Kortix Cloud starts at $40 per seat per month plus usage. The three costs come apart cleanly, and a ten-person team prices out differently on each model.
What actually drives the bill for an AI agent platform
Every AI agent platform bill is made of at most three charges: a seat or licence fee, a model-token bill, and the infrastructure the agents run on. An AI agent platform is the software layer a company runs its agents on, bundling the six layers of an open source AI agent platform: an agent harness, isolated machines, memory, connectors, permissions and a review gate. Closed platforms bundle the first two charges into a subscription and hide the third. Open source platforms invert that by making the software free and leaving you the infrastructure and the tokens.
The seat or licence fee is the recurring charge per person. A closed platform prices it per person and per month, so the bill grows with headcount even when the software does the same work for everyone. A self-hosted open source platform charges no seat fee, so the tenth and the hundredth teammate add nothing to the licence line.
The model-token bill is what the language model charges for the work. Agents read a large context and write less, so input tokens usually dominate. Rates are public and quoted per million tokens: Claude's API rates list Sonnet at $2 per million input tokens and $10 per million output tokens, with Haiku at $1 and $5. When you bring your own key, this bill goes to your provider account and no platform adds a markup.
The infrastructure bill is the compute the agents run on. Self-hosting runs the platform from Docker images on a VPS, a VPC or a laptop, and a suitable server starts at around $24 a month for 4 GiB of memory and 2 vCPUs on DigitalOcean. Managed platforms fold this into the subscription and charge it back through credits or usage.
Two ways to pay: closed managed against self-hosted open source
| Cost driver | Closed managed platform | Self-hosted open source (Kortix) |
|---|---|---|
| Software licence | Included in the subscription | Free to self-host, open source |
| Open source | No | Yes (Elastic License 2.0) |
| Seat fee | Charged per person, per month | None, extra users add no seat cost |
| Model tokens | Vendor models at vendor rates | Any model, your own keys, provider rates |
| Infrastructure | Vendor's cloud, included | Your VPS, VPC or on-prem |
| Self-hosting | Not offered | The same config runs either way |
| Ownership | Configuration in the vendor's product | Configuration is a git repo you own |
A closed platform prices seats like a SaaS product and caps usage per plan, so hitting the cap means a bigger seat tier or metered overage. Claude's plans show the pattern: Pro at $17 a month billed annually, Team at $20 per standard seat a month billed annually, and Enterprise at $20 per seat a month plus API-rate usage. Self-hosted open source removes the seat line: the licence is free, the only fixed cost is the server, and the model tokens are bought at the provider's published rates.
What Kortix costs, as published
Kortix publishes three managed-cloud plans and keeps self-hosting free. As listed on the Kortix pricing page in October 2026:
- Free: $0, with 200 credits a month for sandbox compute, one project, and bring-your-own-key access to any premium model.
- Team: $40 per seat per month, with 2,500 credits a month per seat pooled across the team, about 125 default Agent Computer hours when used only for compute, and up to 200 projects and 100 seats. Bring-your-own-key and a ChatGPT subscription still work, and managed models draw on the same pooled credits.
- Enterprise: custom pricing, adding SAML SSO with SCIM, advanced RBAC and audit logs, plus cloud, VPC or on-prem deployment.
Self-hosting is free. Kortix is open source (Elastic License 2.0) — self-host, read and modify the code. The source is public at Kortix on GitHub. A self-hosted install runs from Docker images as one Docker Compose stack, brought up with kortix self-host start, and the self-hosting guide covers a laptop, a VPS, a VPC or an on-prem network. The same kortix.yaml and the same agents, skills, memory and connectors run on your own host or on Kortix Cloud, so a team switches between them by changing the host.
Worked example: a ten-person team
The numbers assume a ten-person team for one month, at list prices checked in October 2026. The closed column uses Claude's Team plan as the representative per-seat product.
| Line item | Kortix Cloud (Team) | Kortix self-hosted | Closed managed (Claude Team) |
|---|---|---|---|
| Seats or licence | $400 (10 seats at $40) | $0 | $200 to $250 (10 standard seats) |
| Included compute | 25,000 pooled credits, about 1,250 hours | Your server, your hours | Plan usage ceiling |
| Model tokens | Your own keys, billed by provider | Your keys at provider rates | Anthropic models only, no BYOK |
| Infrastructure | Included in the seat price | $24 to $48 VPS a month | Included in the seat price |
| One-month floor | $400 plus tokens | $24 to $48 plus tokens | $200 to $250 plus overage |
Ten Kortix Cloud seats come to $400 a month and include 25,000 pooled credits, about 1,250 Agent Computer hours when used only for compute. Self-hosting replaces the $400 seat line with a rented server, a 4 GiB, 2 vCPU VPS at around $24 a month or an 8 GiB, 4 vCPU machine at around $48, plus tokens at the provider's rates. Team size does not enter the fixed cost.
On the closed plan, ten standard seats are $200 a month billed annually or $250 billed monthly, and the plan caps usage. Exceeding it means premium seats at $100 each a month, Enterprise at a seat price plus API-rate usage, or metered overage. None of those paths lets the team self-host, and the token bill stays on the vendor's model with no bring-your-own-key option.
How to choose between managed and self-hosted
Choose managed Kortix Cloud for a small team that wants no operations work and a start today. Choose self-hosting when the data must stay on hardware you control, when the team is large enough that per-seat pricing would dominate, or when local models run on your own GPUs. The platform and the config are the same, so the decision is reversible with the CLI.
The break-even is simple. Ten seats at $40 pay $400 a month before usage, roughly one to two years of a mid-sized VPS. Below that scale, managed cloud is usually cheaper once an engineer's hours are priced against the ops work. Past 25 seats, the self-hosted seat fee of zero compounds while a managed bill keeps climbing. The best open source AI agent platforms comparison covers how the alternatives price their hosted tiers.
Where self-hosting saves money, and where it does not
Self-hosting saves on seats, on licence renewal and on token markup. The model bill does not disappear, and on an agent-heavy workload it can be the largest line. Self-hosting also moves operational work to your team: upgrades, backups, secrets and the machine image sessions boot. Running local models instead of an API provider adds the capital cost of a GPU to the server line.
FAQ
Do I pay per seat on a self-hosted open source agent platform?+
No. A self-hosted open source platform has no vendor subscription to meter, so there is no seat fee. Kortix self-hosts free, and you pay for the server and the model tokens agents spend.
Is there a markup on model tokens when I bring my own API key?+
No. When you bring your own key, the model provider bills you directly at its published rates and Kortix adds nothing. Claude's rates are $2 per million input and $10 per million output tokens for Sonnet. The Kortix Gateway and managed models are the exception, drawing on plan credits instead of your key.
What does self-hosting an agent platform actually require?+
A machine that runs Docker. Kortix self-hosts from Docker images as one Docker Compose stack, so a 4 GiB VPS is a workable start for a small team. The self-hosting guide walks the prerequisites and install.
How does cost change as the team grows from ten to a hundred?+
Self-hosted, the fixed cost barely moves because one server carries many users and the marginal seat cost is zero. Managed, the seat line scales linearly, so a hundred Team seats list at $4,000 a month before usage. Infrastructure grows with concurrent sessions, while the managed seat line grows with headcount.
Can a team start on managed cloud and move to self-hosting later?+
Yes. The same project and the same kortix.yaml run on both, and the CLI switches hosts with kortix hosts use selfhost and kortix hosts use cloud. A team can start on Kortix Cloud to skip setup, then move to its own VPC or on-prem network when cost or data-residency rules justify it. The AI agent management guide covers roles, secrets and audit, and the platform FAQ answers the rest.
Run the open source AI Management System on your own hardware or on managed cloud: Get started with open-source Kortix.
Keep reading
Open Source AI Agent Platform Comparison Chart and Self-Host an Open-Source AI Agent Platform with Kortix.