Models
Stages that call an AI service need a model. You choose one in the inspector's Model section, or set a default once for the whole channel or blueprint.
If you haven't added any keys yet, the Fake (test) provider is always there: it makes placeholder results for free, so you can build and try a blueprint before paying for anything. See AI provider keys to add real ones.
Pick a model
- Open the stage and find Model.
- Under Provider, choose a provider. This clears any model you'd chosen.
- Under Model, choose a model. The list only shows models that can do this stage's kind of work, such as text or video.
- Adjust the settings below if you need to.
To go back to the default, select Unset model. The stage then uses the blueprint's or channel's default for its kind of work.
Reelcraft doesn't warn you when a stage that calls an AI service has no model and there's no default
for its kind of work. The blueprint still shows Runnable, but when the run reaches that stage it
fails with ProviderRegistry: unknown provider "undefined", after a few retries. If you see that
message, set a model on the stage or a default.
If a provider has no available model for the stage, the box says why, for example that Codex needs its image extension or that BrowserOS Neo isn't running.
Providers
| Provider | What it does in Reelcraft |
|---|---|
| OpenRouter | Text and image models from many vendors, paid by use with your OpenRouter key |
| fal | Video generation |
| ElevenLabs | Speech (voice-over), with its own voices and settings |
| Deepgram | Speech (voice-over), and transcription with word timings for Analyze Media |
| Codex | Text, and image and browser steps, using your ChatGPT plan. See Connect Codex |
| Claude | Text and browser steps (no images), using your Claude plan. See Connect Claude |
| ChatGPT | Text and images through a ChatGPT tab in BrowserOS Neo. See BrowserOS Neo |
| Fake (test) | Free placeholder results for every kind of work |
Which providers appear depends on what you've set up in Settings.
Model settings
Generate Speech has its own Model section: model cards, a voice picker with samples and a form of the settings that model supports. The rest of this section is for the other stages.
Which of these you see depends on the provider.
- Version: pin a specific version or snapshot of the model, if the provider has them. Leave it empty to use the provider's current one.
- Effort (Codex, Claude and ChatGPT): how hard the model thinks before answering. Higher is slower and uses more of your plan. The choices depend on the model.
- Web search (ChatGPT): let the model search the web for this stage's prompt.
- Params: extra settings passed to the provider on every call, such as
max_tokensortemperature. Select + add param, name it, and enter a value as text, a number or true/false. Reelcraft sends them to the provider as you typed them, so use names the provider documents. The same params also go to this stage's quality control call.
Default or per stage?
A stage's own model always wins. Otherwise Reelcraft uses the blueprint's default for the stage's kind of work, then the channel's. See Defaults.
A good pattern is to set text, image, video and speech defaults on the channel, and then only pin a model on a stage that needs a different one, such as a more capable model for a tricky script.
Models and cost
Each provider bills in its own way, and Reelcraft estimates the cost of a call before making it. See Budget and costs and Costs.
The quality control judge
A stage's quality control has its own model, chosen in the same way but separately: it doesn't use these defaults.