Provider integration guide

Concepts

A provider is the connection configuration for one external service. Each type declares capability tags (llm / tts / asr / image / video); calls are routed by tag, not by vendor.

Built-in types

Type	Capability	Description	Default base_url
openai_compatible	llm	vLLM / LM Studio / DeepSeek / Qwen / Volcano Ark / Kimi / Zhipu / SiliconFlow / OpenRouter	http://localhost:8000/v1
ollama	llm	Local Ollama (models from /api/tags, chat via /v1)	http://localhost:11434
openai_tts	tts	OpenAI / Kokoro-FastAPI / Speaches / SiliconFlow	http://localhost:8880/v1
edge_tts	tts	Microsoft online TTS, zero-config fallback (mp3 only)	—
openai_stt	asr	OpenAI / Groq / Speaches / faster-whisper-server	http://localhost:8100/v1
comfyui	image, video	LAN ComfyUI, driven by workflow packs	http://localhost:8188
openai_images	image	OpenAI gpt-image and compatible gateways	https://api.openai.com/v1

Add a provider

In the Web UI under Settings → Providers, or with the CLI:

astra provider add --name home-ollama --type ollama --base-url http://192.168.1.10:11434
astra provider test 1
astra provider models 1

• api_key is Fernet-encrypted before storage; the API only returns has_api_key.
• concurrency limits concurrent jobs for that provider (default 2).
• Extra configuration (such as a default model) goes in extra: {"model": "qwen3:8b"}.

ComfyUI workflow packs

A ComfyUI provider is driven by workflow packs. A pack is a workflow.json (ComfyUI → Save (API Format)) plus a manifest.yaml:

name: my_t2i
capability: t2i            # canonical: t2i / i2i / upscale / t2v / i2v / avatar / motion / tts
inputs:
  prompt: { node: "6", field: inputs.text, type: str, required: true }
  image:  { node: "1", field: inputs.image, type: image }   # image/video/audio inputs are uploaded to ComfyUI automatically
  seed:   { node: "3", field: inputs.seed, type: int, default: -1 }  # -1 = random
outputs: { node: "9", key: images }   # images / videos / gifs
# Optional: acceleration LoRA (removed when disabled; downstream reconnects to bypass_input)
# accel:
#   node: "200"
#   bypass_input: model
#   lora: { field: inputs.lora_name, options_from: loras, match: "minimax_h3", recommended: "xxx.safetensors" }
#   strength: { field: inputs.strength_model, default: 1.0 }
#   preset: { steps: 8 }
requirements:
  nodes: [KSampler]        # used by import preflight

Import options: Web UI Settings → Workflow packs (LiberAstra packs, API Format, and canvas JSON with nodes/links converted automatically), or write to data/workflows/<name>/ (a user pack with the same name overrides the built-in one). Video examples: built-in packs minimax_h3_i2v, minimax_h3_i2v_uncensored, minimax_h3_i2v_enhancer.

Add a provider type (development)

Register capability tags in backend/liberastra/providers following the existing provider types, and add the type to the capability registry. The Web UI Providers page uses capability tags to decide which pages and workflows are available.