tongflow-api-openai

Official TongFlow plugin. Text and image generation via any OpenAI-compatible endpoint — OpenAI itself, Azure OpenAI, a third-party relay, or a local server (vLLM / Ollama / LM Studio). Point it at an endpoint with OPENAI_BASE_URL and supply that endpoint's key and models.

Capabilities

Implements these ABI slots (runs locally as a Python process, no GPU):

Text — via the Chat Completions API (/chat/completions):

• Generate / rewrite text (gen-text) — create or edit copy from a prompt.
• Split long text (split-text) — break a long passage into chunks.
• Arrange & batch groups (arrange-group) — group and arrange text/clip batches for downstream processing.
• Filter or drop clips (drop-video) — drop unwanted clips by rule.

Image — via the Images API (/images/generations, /images/edits), using gpt-image-2 by default:

• Generate image (image-gen) — text → image.
• Edit image (image-edit) — one image + an instruction → edited image.
• Fuse images (image-fusion) — multiple reference images + a prompt → composed image.

Credentials

Add in TongFlow Settings (gear icon, top-right):

Key	Required	Notes
OPENAI_API_KEY	✅	Key for the target endpoint. For OpenAI itself, create one at platform.openai.com/api-keys.
OPENAI_BASE_URL	optional	OpenAI-compatible endpoint. Defaults to https://api.openai.com/v1; set it to an Azure / relay / local (vLLM·Ollama·LM Studio) base URL to switch providers.
OPENAI_CHAT_MODEL	optional	Chat model id for the text slots (default gpt-4o-mini).
OPENAI_IMAGE_MODEL	optional	Image model id for the image slots (default gpt-image-2; e.g. gpt-image-1, dall-e-3).
OPENAI_IMAGE_SIZE	optional	Force an output size for the image slots, e.g. 1024x1024 / 1536x1024 / auto. By default the node's width × height is used, or the model's default when unset.

Values are stored locally and take effect without a restart.

OPENAI_BASE_URL must be the API root, not the full path. The plugin appends /chat/completions, /images/generations, or /images/edits itself, so set the base URL up to the version segment (e.g. https://api.openai.com/v1) — do not include a method path. A trailing slash is fine; it is trimmed automatically.

Image slots require an endpoint that serves the OpenAI Images API. Most relays and Azure expose it; local servers (Ollama / vLLM / LM Studio) typically do not — use them for the text slots only.

Pointing at a different provider

Any service that implements the OpenAI Chat Completions API works. Set the three variables to match the provider:

Provider	OPENAI_BASE_URL	OPENAI_API_KEY	OPENAI_CHAT_MODEL
OpenAI (default)	(leave empty)	your OpenAI key	e.g. gpt-4o-mini
Third-party relay / proxy	the relay's base URL, e.g. https://your-relay.example.com/v1	the relay's key	a model the relay exposes
Azure OpenAI	https://<resource>.openai.azure.com/openai/v1	your Azure key	your deployment name
Local — Ollama	http://localhost:11434/v1	any non-empty string	e.g. llama3.1
Local — vLLM / LM Studio	http://localhost:8000/v1 (vLLM) · http://localhost:1234/v1 (LM Studio)	any non-empty string	the served model id

Mixing an endpoint with a key issued for a different endpoint is the common failure: a relay key sent to api.openai.com returns 401 Unauthorized. Make sure OPENAI_API_KEY belongs to whatever OPENAI_BASE_URL points at.

Errors

Failures surface the endpoint that was actually called plus the upstream response body, e.g. HTTP 401 from OpenAI (https://.../v1/chat/completions): {...}. Use it to tell apart a wrong key (401), rate limiting / quota (429), and a wrong base URL (404 / connection error).