Menu

Mattering Automations Together AI integration

Connect Together AI to Mattering Automations: chat and text completions on open-weight models, embeddings, rerank, FLUX image generation, Whisper audio, batch jobs, files, and fine-tuning as workflow steps, authenticated with a Together AI API key.

The Mattering Automations Together AI integration puts Together's open-weight model inference on your automation canvas: run chat or text completions on Llama, Qwen, or DeepSeek inside a workflow step, generate FLUX images that come back as hosted URLs, create embeddings and rerank documents, transcribe or translate audio from a URL with Whisper, submit and poll discounted batch jobs, manage the files those jobs read and write, and run a fine-tuning job from creation through to a callable model.

Setup is one API key from your Together account. This guide covers creating the key, connecting it, a verify path that costs nothing to run, and the full node reference, and it ends with the failure modes worth recognizing on sight.

Before you start

  • A Together AI account at api.together.ai. New accounts come with trial credit, so a key works for testing before any payment is set up.
  • A Mattering workspace with access to the Automations app.
  • For the audio nodes, somewhere to host audio at a publicly reachable URL. Together fetches the audio itself; there is no local file upload.
  • For batch and fine-tuning workflows, the JSONL input file uploaded out of band with Together's CLI or SDK. There is no upload node, but everything from the file id onward runs in the workflow.
  • For the rerank node, a dedicated endpoint serving a rerank model. Together currently serves rerank models on dedicated endpoints rather than serverless.
  • For fine-tuning, a clear-eyed look at Together's pricing first: creating a job launches paid GPU training billed by trained tokens.

Create a Together AI API key

Mattering authenticates to Together with an API key sent as an Authorization bearer header. It is a single field: no organization id, no version header, no OAuth flow to complete.

Sign in at api.together.ai and open Settings, then API Keys.
Create a new key and name it after where it will be used, such as Mattering Automations.
Treat the key like a password: it can run inference, start paid fine-tuning jobs, and delete files on the account, all billed to you.
Copy the key and keep it somewhere safe until you have pasted it into Mattering.

Connect Together AI to Mattering

In Mattering Automations, open Connectors and click Add Connector.
Choose Together AI from the list.
Click Create new credential, paste the API key, and save.
The credential is stored encrypted in Google Secret Manager and never appears in workflow logs.
Click Connect. All 22 Together AI nodes appear in your canvas palette.
The Together AI connector does not ship a built-in auth test, so Test Connection has nothing to call for this one. Verify with the read-only node in the next section instead.

Verify it works

Open any automation canvas and check the palette: a Together AI section should list nine query nodes and 13 action nodes. Drop Together AI: List Models onto the canvas and run a test step. It takes no parameters and is read-only, so a successful response returns every model your key can call, each with its type, context length, and per-token pricing. It costs no tokens, which makes it the cheapest credential check available.

To confirm inference itself works, add Together AI: Create Chat Completion, set the model to a small current model id from that list, put one short message in the messages array, and keep max_tokens low. A successful run returns the reply at choices[0].message.content along with token usage.

Node reference

22 nodes, generated from the connector itself so this reference always matches what ships.

Read and query nodes 9

Together AI: Download File Content

together_get_file_content

Download a file's contents as text. Batch output and error files are JSONL, so the node also returns each line parsed into a lines array you can loop over.

Together AI: Get Batch

together_get_batch

Get one batch job: status, progress percentage, and the output and error file IDs to download once it completes.

Together AI: Get File

together_get_file

Get one file's metadata: size in bytes, filename, purpose, and processing status.

Together AI: Get Fine-Tuning Job

together_get_fine_tuning

Get one fine-tuning job: status, epochs completed, token count, and the model_output_name you pass as model on a chat completion once training finishes.

Together AI: List Batches

together_list_batches

List the batch jobs on the account with status (VALIDATING, IN_PROGRESS, COMPLETED, FAILED, EXPIRED, CANCELLED) and progress.

Together AI: List Files

together_list_files

List the files uploaded to the account, with size, filename, purpose, and processing status.

Together AI: List Fine-Tuning Events

together_list_fine_tuning_events

List the event log for a fine-tuning job: pending, training start, checkpoint saves, completion, or errors. Useful for polling training progress from a flow.

Together AI: List Fine-Tuning Jobs

together_list_fine_tunings

List the fine-tuning jobs on the account with status, base model, and output model name.

Together AI: List Models

together_list_models

List every model available to the API key, with type (chat, language, image, embedding, rerank, moderation), context length, and per-token pricing. Works with no parameters, so it doubles as the credential health check. Together documents no single-model GET, so this list is the lookup surface.

Action nodes 13

Together AI: Cancel Batch

together_cancel_batch

Cancel a batch job that is still validating or in progress.

Together AI: Cancel Fine-Tuning Job

together_cancel_fine_tuning

Cancel a running fine-tuning job. Billing stops for work not yet performed.

Together AI: Create Batch

together_create_batch

Kick off an asynchronous batch job over a previously uploaded JSONL file. Batch runs are discounted and write results to an output file when they finish. The response wraps the job under data.job.

Together AI: Create Chat Completion

together_create_chat_completion

Run a chat completion on a Together-hosted open-weight model (Llama, Qwen, DeepSeek, Mixtral, and more). Send a messages array, get the assistant reply at choices[0].message.content plus token usage. Supports tool calling and JSON-mode structured output. Non-streaming.

Together AI: Create Completion

together_create_completion

Run a raw text completion: send a prompt string, get the continuation at choices[0].text. Use this for base language models or when you want to control the prompt template yourself; use Create Chat Completion for instruct models. Non-streaming.

Together AI: Create Embedding

together_create_embedding

Turn text into embedding vectors for search, clustering, or RAG. Send one string or a JSON array of strings; each input comes back as a float vector at data[n].embedding.

Together AI: Create Fine-Tuning Job

together_create_fine_tuning

Start a fine-tuning job (LoRA by default, full fine-tuning optional) on an uploaded JSONL training file. This launches paid GPU training billed by token count, so check pricing before wiring it into an automation. The training file must be uploaded out-of-band (CLI or SDK) first.

Together AI: Delete File

together_delete_file

Permanently delete an uploaded file from the Together account.

Together AI: Delete Fine-Tuning Job

together_delete_fine_tuning

Delete a fine-tuning job record from the account. This cannot be undone.

Together AI: Generate Image (URL)

together_create_image

Generate an image from a text prompt with FLUX or another Together-hosted image model. The result comes back as a hosted image URL at data[0].url, never inline base64, so pipe the URL onward or download it promptly if you need to keep the image.

Together AI: Rerank Documents

together_create_rerank

Rank a list of documents by relevance to a query. Returns each document's index and relevance_score, best first. Note: Together currently serves rerank models on dedicated endpoints rather than serverless, so check List Models for what your account can call.

Together AI: Transcribe Audio (URL)

together_create_transcription

Transcribe an audio file into text with Whisper on Together. Takes a publicly reachable audio URL (Together fetches it, up to 1 GB); this node does not upload local files. Supports speaker diarization. Formats: wav, mp3, m4a, webm, flac, ogg, opus, aac.

Together AI: Translate Audio to English (URL)

together_create_translation

Translate non-English audio into English text with Whisper on Together. Takes a publicly reachable audio URL (up to 1 GB); this node does not upload local files.

Using the integration

On the automation canvas

Together AI nodes chain with everything else: pull a record, generate or classify with Together AI: Create Chat Completion, then write the result into a task, a message, or a database row. Template references resolve inside nested JSON, so you can drop an upstream node's output straight into one entry of the messages array, the documents list on a rerank, or a tools definition, rather than templating the whole array as a single blob.

Chat completions or raw text completions

Together AI: Create Chat Completion is the familiar messages in, choices out shape, with tool calling and JSON-mode structured output, plus Together-specific controls that OpenAI does not have: top_k, min_p, repetition_penalty, a safety_model field that runs a Llama-Guard moderation pass, and context_length_exceeded_behavior, which can truncate instead of erroring when a prompt overruns the window. Together AI: Create Completion sends a raw prompt string for base language models or when you want to control the prompt template yourself. Both are non-streaming, so each node returns one complete response.

Images that arrive as URLs

Together AI: Generate Image (URL) renders a prompt with FLUX or another Together-hosted image model and returns a hosted link at data[0].url. The connector pins the response format to URL on purpose: no base64 ever lands in an execution record, so image-heavy workflows stay light, and the URL pipes straight into the next node. Together does not document how long the hosted URL stays live, so add a download step into your own storage whenever the image matters beyond the run.

Embeddings and rerank for search

Together AI: Create Embedding takes one string or a JSON array of strings and returns a float vector per input at data[n].embedding, ready for search, clustering, or RAG. Together AI: Rerank Documents scores a candidate list against a query and returns each document's index and relevance_score, best first. Rerank models currently require a dedicated endpoint, so confirm availability with Together AI: List Models before building on it.

Audio in, text out

Together AI: Transcribe Audio (URL) and Together AI: Translate Audio to English (URL) take a publicly reachable audio URL that Together fetches for itself, up to 1 GB. Accepted formats are wav, mp3, m4a, webm, flac, ogg, opus, and aac, and transcription supports speaker diarization when you need to know who said what.

Batches and the files around them

Upload the JSONL input out of band, then Together AI: Create Batch starts a discounted asynchronous run over that file id and returns the job under data.job. Together AI: Get Batch reports status and progress plus the output and error file ids once it completes, and Together AI: Download File Content returns the JSONL as text with each line already parsed into a lines array a For Each node can loop. Together AI: Cancel Batch stops a run that is still validating or in progress, and Together AI: Delete File cleans up what a finished batch left behind.

Fine-tuning as a lifecycle, not a black box

Together AI: Create Fine-Tuning Job starts LoRA training by default, with full fine-tuning optional, on an uploaded training file. This is the one node in the connector that launches paid GPU training, billed by trained tokens, so treat it deliberately. From there the job is fully observable from a flow: Together AI: List Fine-Tuning Events polls the event log for training start, checkpoint saves, and completion, Together AI: Cancel Fine-Tuning Job stops billing for work not yet performed, and the finished job's model_output_name goes straight into the model field on a chat completion.

From AI assistants over MCP

Every node speaks MCP, the open protocol for connecting AI assistants to tools. Claude, ChatGPT, and Matter can drive them, so an assistant can build and run a workflow that hands the generation, the images, or the bulk batch to a Together-hosted open model.

Troubleshooting

Every node returns a 401
The key is wrong, revoked, or truncated on paste. Create a fresh key at api.together.ai under Settings, then API Keys, update the credential on the connector, and run Together AI: List Models to confirm the fix. It is read-only and costs no tokens.
The rerank node returns a model availability error
Together currently serves rerank models on dedicated endpoints rather than serverless, so Together AI: Rerank Documents fails on an account with no dedicated rerank endpoint no matter how the node is configured. Run Together AI: List Models with its dedicated filter to see what your account can actually call. If nothing rerank-shaped appears, the fix is on the Together side: provision a dedicated endpoint for a rerank model.
A generated image URL no longer loads
Image responses are hosted URLs by design, and Together does not document how long they stay live, so a URL saved for later can stop resolving. Do not treat the hosted URL as permanent storage. Put a download step right after Together AI: Generate Image (URL) that writes the image into storage you control, and reference your own copy from then on.
A fine-tuning job will not start, or you are unsure about cost
First check the training file: the id must come from a file already uploaded with Together's CLI or SDK, and Together AI: Get File should show it fully processed before a job can train on it. A job can also be rejected when the account has hit Together's fine-tuning limits or lacks the credit to cover training, because Together AI: Create Fine-Tuning Job launches paid GPU training billed by trained tokens; the job object reports the price. A running job you regret can be stopped with Together AI: Cancel Fine-Tuning Job, which stops billing for work not yet performed.
A node returns a model not found error
The model id is not available to your key, or Together has retired it from serverless. The catalog changes on a rolling basis, so an id copied from an old workflow can stop resolving. Run Together AI: List Models to see the current set and paste one of those ids into the model field. The field is free text on purpose, so it does not rot the way a fixed dropdown would, and Together documents no single-model lookup, so the list is the check.
Create Batch returns a job instead of results
Batches are asynchronous by design. Together AI: Create Batch returns the new job under data.job, not the completions. Poll with Together AI: Get Batch on a later scheduled run until the status reads COMPLETED, note that batch statuses are uppercase, then read the results with Together AI: Download File Content using the output file id the batch reports.
An audio node fails because Together cannot fetch the recording
Transcription and translation hand Together a URL and Together downloads the audio itself, so a link that only resolves inside your network, an expired signed URL, or a page behind a login all fail on Together's side rather than in the workflow. Open the URL in a private browser window with no session: if the audio downloads there, Together can reach it too. Check the file is under 1 GB and in one of the accepted formats: wav, mp3, m4a, webm, flac, ogg, opus, or aac.
There is no node for the endpoint you need
Some gaps are deliberate. File upload needs a binary multipart payload the canvas config panel cannot produce, so upload the JSONL or training file out of band and bring the file id into the workflow; the list, metadata, download, and delete nodes cover the rest of the lifecycle. Text to speech returns raw audio bytes rather than JSON and is not shipped. Streaming is not exposed, and the connector strips stream from outbound bodies so a pasted OpenAI payload cannot silently break a node. Together's video generation and code execution products are not covered today, and Together itself documents no single-model lookup, so no node can exist for that.

Together AI is a trademark of its respective owner. Mattering uses platform names to identify the services these integrations connect to; no endorsement or affiliation is implied.

Start Free Explore our Products