DocsOverview

VidSwarm V1 and agent profiles

VidSwarm's AI comes in two presets, Lite and Pro. An agent profile says where it runs and which model does each step.

Lite and Pro

VidSwarm V1 is the set of models behind every agent: the video agent (research, script, package, translation, your edits), the ideation agent, the thumbnail agent, vision and the voice. A preset fills every step with the right kind of model.

  • Lite: fast, low-cost models such as DeepSeek and Claude Haiku. The default, and what the free plan runs.
  • Pro: Claude Sonnet class for research, script, titles and your edits; stronger writing, at the Standard rate.

The credits per finished minute for each are on the Pricing page and in Credits and plans. Stronger models (Claude Opus class) can be picked for single steps.

Where it runs

On a plan, Settings › AI models has a Run on choice: VidSwarm cloud (credits), your Claude, ChatGPT/Codex or Gemini subscription through the desktop app, or your own API key. Each step then uses the closest model there; steps on your subscription or key cost 0 AI credits, and storage, checks and publishing stay.

VidSwarm V1 is tuned and tested on Claude models. On GPT or Gemini each step uses the closest model and the page says that quality may differ. Steps a place cannot do (the desktop app's subscriptions do not look at frames) stay on VidSwarm cloud, and the overview says why.

Your own subscriptions and keys run at your own risk, under each provider's terms. The first time a profile uses one, an owner confirms that once for the workspace; we store who and when.

The Gemini subscription runs through Google's Antigravity CLI in the desktop app (Gemini CLI only works with a Gemini API key now). Google's terms don't allow third-party apps to use Gemini subscriptions, so a profile on it keeps a warning line. A Gemini API key has no such limit.

Agent profiles

Who does what shows each agent's steps with the model, where it runs and what it costs. Change any step to another model and source; Back to preset undoes it. Save the result as a profile, for example "Local via my Claude subscription", and make one the default.

The default runs everywhere. A channel can use another profile (Channel settings › General › AI profile), and a film can pick one when it starts (the voice step). Fallback models and what happens when every model fails are part of the profile.