App-store UX in the desktop
The model library is a first-class surface — not buried in advanced settings. Each catalog tile shows a clear primary action (install, download, or use). Open a detail page when you need the full flow: progress, delete, import to Ollama, or assign to a specific agent.
Ollama catalog
- Curated list via
GET /api/ollama/catalogembedded with the hub. - Install with streaming progress; remove models you no longer need.
- Use for agents sets the Ollama provider model and agent assignments in one step.
Hugging Face catalog
- Hosted — route agents through the HF Inference Router when you have
HF_TOKENor a provider key. - Download — pull GGUF artifacts or LoRA adapter safetensors with SSE progress; Import to Ollama or Compose & import for local runs (see LoRA adapters).
- Create DM experts against a registry provider or a
huggingfacerepo id from the same UI.
Local image generation
Beyond chat models, Ollama can run local image generation. Configure under Settings → Image generation; agents use hub-side tools to create and attach images in chat.
Where settings still live
Settings → AI Providers is for the provider registry, API keys, and Ollama endpoint — not day-to-day model shopping. Day-to-day install/browse happens in the Model library toolbar entry.
Mix local and cloud
Keep high-volume agents on local Ollama (qwen2.5-coder, Llama, vision-capable tags where listed) and burst review-heavy personas to Claude or OpenAI-compatible APIs. Per-agent provider routing stays independent of the catalog you picked in the store.