Pricing

One rule decides every price here: if the work runs on your own machine, it is free, because it costs us nothing. If it runs on ours and we are billed for it, you are charged what it costs plus a stated margin. Which side a thing falls on is a property of the work itself, not a plan you are on.

Free, and not free-for-now

These run with no API key and no outbound call, so there is nothing for us to bill you for: searching your codebase, asking it a question in plain language, comparing two versions of a project, rendering a scene to an image or a PDF, and reading or publishing knowledge entries. The embeddings behind the search are local by design — the factory that chooses them is forbidden from silently picking a paid provider.

Five of those carried a price until 2026-09-21 and should not have. They are zero now.

Charged, and why

Five things call a paid model on someone else's hardware: code autocomplete, code generation, chat, material generation, and turning speech into HoloScript. Those cost us money on every request, and we charge you nothing for them. That is a decision, not an oversight, and it is the answer to what the free tier is: the parts where we pay a third party on your behalf.

This page said the opposite until 2026-09-21. It listed those five as charged and said we could account for them line by line. Nothing collected them and nothing ever had.

Every price still on this page is under review

11 operations carry a price today, and we audited all 11 on 2026-09-21 without being able to justify a single one on the spot. 7 are genuinely collected but we have not traced them to a cost we actually pay. 4 are listed while nothing collects them at all. Treat both kinds as under review rather than as a quote: where the work turns out to run on your own hardware the price goes to zero, as five already did.

We would rather publish that sentence than a price we cannot explain. It is the second correction this page has carried in a day; the first one read better and was false.

Credits

One credit is one cent. Buying more at once costs less per credit; the larger packages include bonus credits, so what you pay and what you receive are two different numbers.

Model tokens

When one of the five charged operations calls a paid model, the tokens are metered at what the provider charges us plus a 15% margin. Run a model on your own hardware or on your own fleet instead and the token cost is zero — that lane is free, and it is the one we would rather you used.

Or pay nothing at all

The language, the command-line tool, the engine, the codebase intelligence, the embeddings and the mesh are MIT licensed. You can run the whole thing on your own hardware and never pay us anything. We think that is the right arrangement: we should be paid for compute we buy on your behalf, not for access to a language.