Together AI integration

What it does

Together AI hosts hundreds of open-source and partner models behind one API. Connect a Together AI account and your workflows can write replies and structured JSON with Llama, DeepSeek, Qwen or GLM, generate images with FLUX and similar models, turn text into speech, transcribe call recordings with speaker labels, translate spoken audio into English, and generate short videos. Every model field is a picker backed by the connected account's live model list, so you choose from models the account can actually call instead of pasting IDs.

Connect a Together AI account

Together AI authenticates with a single API key.

1. Sign in at api.together.ai. New accounts get a small starting credit and no card is needed to begin.
2. Open Settings, API keys, create a key, and copy it.
3. In TaskJuice, add a Together AI node to a workflow, open its connection picker, and paste the key.

A key reaches every model and endpoint the account can use. To rotate it, create a new key, update the connection, then delete the old key in Together AI.

Triggers

Together AI publishes no webhook or event-subscription API, so this integration ships actions only. Put a Together AI action downstream of whichever trigger starts the work, such as a form submission, a new CRM record or a schedule.

Actions

Text

• togetherai/create-chat-completion: generate a chat reply from a chat model, with optional JSON output, temperature, max tokens and stop sequences.
• togetherai/create-completion: continue a raw text prompt with a language model.
• togetherai/create-embedding: turn text into an embedding vector for search or retrieval.

Images

• togetherai/generate-image: create one or more images from a prompt, returned as hosted URLs or base64 data.

Audio

• togetherai/create-speech: turn text into an MP3, WAV or raw audio file.
• togetherai/list-voices: list the voices each text-to-speech model offers.
• togetherai/transcribe-audio: transcribe an audio file from an earlier step, with optional speaker labels and word timestamps.
• togetherai/transcribe-audio-from-url: transcribe audio at a public link, such as a call recording URL.
• togetherai/translate-audio: transcribe an audio file and translate the speech into English.

Video

• togetherai/create-video: start generating a video from a prompt. Returns a job ID straight away.
• togetherai/get-video: check a video job and collect the finished video URL.

Models

• togetherai/list-models: list every model the account can call, with type, context length and pricing. The models are under value.

Known limitations

Video generation is asynchronous. Create Video returns while the video is still rendering. Add a Delay of a few minutes, then Get Video with the job ID. When status reads completed, outputs.video_url holds the link; branch on status to handle in_progress (wait longer) and failed (read error.message).

Embedding models run on dedicated endpoints. Together AI currently serves no embedding or rerank models on serverless, so Create Embedding needs a dedicated endpoint for the model you pick. The picker lists the embedding models the account can see.

Voices belong to a model. Each speech model has its own voice names, for example af_alloy for Kokoro or tara for Orpheus. Run List Voices once to see them, and an unknown voice is rejected by Together AI.

Uploaded audio is capped at 80 MB. Transcribe Audio and Translate Audio accept files up to 80 MB and 4 hours. For larger recordings at a public link, use Transcribe Audio from URL, which accepts up to 1 GB.

Image and video links expire. Hosted URLs returned by Generate Image and Get Video are temporary. Save the file to storage in the same workflow if you need it later, or ask Generate Image for base64 data.

Spend limits and context length. A 402 means the account hit its monthly spending limit or ran out of credit. A 403 usually means the input plus Max Tokens exceeds the model's context length. Neither is retried. Rate limits (429) and Together AI server errors are retried with backoff.

Streaming is not exposed. Each call waits for the full response.