Connecting a provider
Window → Customize → AI Engine.
The assistant is in both editions. It is bring-your-own-key either way. The difference is where the key lives: in the installers it is sent to the server and kept encrypted there; in the local edition (built from source, which never contacts our servers) it stays on your computer, encrypted with the operating system’s keystore — Windows’ DPAPI, the macOS Keychain — and every model call goes straight from your machine to the provider.
Two ways to power it
Your own key works in both editions. Motion AI needs the server, so it is a server-edition option only.
| Option | Status |
|---|---|
| Motion AI — the first-party provider account, metered and billed, nothing to manage | Disabled while in development. The server says so if you try |
| Your own key — OpenAI, Anthropic or Gemini | Works today |
How your key is handled
You paste it once
In Window → Customize → AI Engine, choose a provider and paste the key.
It is stored encrypted
In the installers, the key is sent once to the backend gateway and stored AES-encrypted against your account, server-side. In the local edition, it is encrypted by the desktop shell with the operating system’s keystore and kept in a file that cannot be read on another machine or under another user account.
The interface never keeps it
It is not held in the interface process and not written to browser storage. In the local edition there is no channel through which the interface can read the key back at all: it can ask for a model call to be made, but it cannot ask for the key.
Streaming is proxied
In the installers, the server’s /ai/stream endpoint proxies the provider’s byte
stream. In the local edition, the desktop shell does. Either way the agent loop
itself runs in the editor, against your live scene.
This design exists for a reason. The interface holds your account token; a provider key sitting beside it in local storage would be one more thing worth stealing.
Choosing a model
The composer has a provider and model picker. Which model to use is a cost-versus-capability trade-off you make per task:
- Structural work — “build a five-card layout, stagger them in” — a smaller model handles well.
- Creative direction — “make this feel like a premium hardware launch” — benefits from a larger one.
Credits and billing
When the first-party provider is enabled, AI usage is metered per plan through the gateway. The plan catalogue, the credit balance and whether checkout is open all come from the server — nothing about billing is decided client-side.
With your own key, you are billed by your provider directly and Premation meters nothing.
If it is not working
| Symptom | Likely cause |
|---|---|
| The panel says the assistant is unavailable | Installer or server edition, but the companion server is not reachable |
| The provider is rejected | The key is wrong, or the account has no quota. The error text comes from the provider |
| Motion AI is greyed out | Expected — it is disabled during development. Use your own key |
| A run starts and nothing lands | Check the mode: Manual waits for you to press Apply |