Providers

Torque routes work to providers through governed model roster slots. Arbiter rules decide which provider handles each task based on type, cost budget, confidence, and whether cloud keys are available.

Local providers

Ollama is the default text generation provider. It runs on your machine — no API key, no network, no per-token cost. Torque communicates with Ollama over its local HTTP API.

ComfyUI is an optional local image generation provider for operators with GPU hardware. Torque submits workflows to a local ComfyUI instance and polls for completed outputs.

Cloud providers

Cloud providers are opt-in. Configure them by setting environment variables for API keys — Torque detects available keys at startup and enables providers accordingly.

OpenAI — text generation (GPT models) and image generation (DALL-E). Used for capable-tier text tasks and storyboard frame generation.

Anthropic — text generation (Claude models). Used for precise-tier and analytical tasks, deep research, and code generation.

OpenRouter — text model routing across multiple providers through a single API.

Replicate — image models (Flux, SDXL) and video models (Minimax Video-01, Hunyuan Video). Used for Cinema storyboard and animatic generation.

Runway — video generation (Gen-4 Turbo). Alternative video provider for animatic generation.

Model roster

The model roster maps abstract capability slots to concrete provider/model pairs. Slots include: fast-local (quick local tasks), capable (quality text generation), precise (high-accuracy work), analytical (reasoning-heavy tasks), image-cloud (image generation), and video-cloud (video generation). Each vertical configures its roster independently.

Cost governance

Every cloud provider call is tracked. Arbiter cost rules enforce daily spend limits, hourly token caps, and per-task budgets. The governance layer throttles or halts execution before overspending — you set the ceiling and Torque respects it.