Model library (toolbar)
Open Model library from the toolbar (⇧⌘M or /nj-open-model-library). Browse Ollama and Hugging Face catalogs in an app-store grid; one primary action per tile; detail pages for full install, use, download, and import flows. Ollama uses GET /api/ollama/catalog with SSE pull progress; Hugging Face adds hosted inference and local GGUF download (GET /api/hf/catalog, POST /api/hf/download, POST /api/hf/import-ollama). Settings → AI Providers keeps the provider registry and Ollama endpoint only.
Hugging Face tab
In the model library Hugging Face tab, a curated catalog (GET /api/hf/catalog) supports Hosted (Inference Router; HF_TOKEN or per-provider key) and Download (GGUF with SSE progress, then Import to Ollama). Create DM experts with a registry provider_id or huggingface + repo id.
Supported provider families
- Ollama — local inference with optional lifecycle help (detect, start/stop, health checks), the model library above, and classic pull/list flows from the hub and desktop.
- Hugging Face — hosted inference via the router API, or local GGUF download + Ollama import (see above).
- Claude — Anthropic API keys via settings or
env.local. - LM Studio — OpenAI-compatible local server presets.
- OpenAI-compatible — Amazon Q, Azure OpenAI, Together, Groq, or any URL that speaks the compatible chat API.
- CLI providers — Cursor and Gemini CLIs when installed, for teams that want those stacks inside the same channel fabric.
Per-agent routing lets you keep cheap local models on high-volume agents while pointing review-heavy personas at larger cloud models.
MCP export and import
Agents accumulate useful context. MCP export serializes agent knowledge into a portable format you can check into git or import into other MCP-aware tools. Import paths round-trip for sharing presets across machines.
Optional MCP tool servers
For certain personas the hub can spin companion MCP servers (gated by configuration) so agents can call structured tools beyond raw chat completion. Toggle behavior via environment flags documented in env.example.
Packaging and updates
The Go hub compiles to a platform-specific sidecar that ships beside the Tauri bundle — one .dmg or .AppImage style artifact without asking users to install Go separately. The Tauri updater channel delivers patch builds once you enable update checks.