The Fireworks AI MCP for open-weight model inference inside your automations
Connect one Fireworks API key and chat completions, embeddings, document reranking, and the fine-tuning lifecycle become steps on a visual canvas that AI agents can drive. 16 typed nodes over Fireworks serverless inference and its control plane.
What the Fireworks AI MCP connector does
The Fireworks AI connector for Mattering Automations covers the useful slice of Fireworks as 16 typed nodes: chat and text completions on serverless open-weight models such as Kimi, Qwen, Llama, and DeepSeek, an Anthropic-format Messages endpoint, embeddings, document reranking, and the control-plane reads that make the rest workable: accounts, the model library, datasets, and supervised fine-tuning jobs. Eight are read-only queries and eight are actions.
Every node speaks MCP, the open protocol for connecting AI assistants to tools. Claude, ChatGPT, and any MCP-capable assistant can call them, and so can Matter, the assistant built into the platform. The shape that pays off here is an assistant planning a workflow and handing the generation, embedding, and reranking work to Fireworks, with everything sitting on one canvas you can see and edit.
Fireworks splits its API into two surfaces on one host: an OpenAI-compatible inference path and a control plane for accounts, models, datasets, and fine-tuning. One detail trips up almost everyone: there is no models listing on the inference surface. Model discovery is the control-plane Fireworks: List Models call, which defaults to the shared public library so it works with zero configuration, and Fireworks: List Accounts hands you your own account ID for everything else. The same API key covers both surfaces, so setup is still one credential.
One connection, three surfaces
- Visual automations. Drag Fireworks AI nodes onto the canvas and wire them to triggers, conditions, and other apps.
- AI agents, yours or ours. These nodes speak MCP, the open protocol, so Claude, ChatGPT, and any MCP-capable assistant can use them, and Matter, the agent built into the platform, uses them natively.
- The whole platform. Combine with native tasks, projects, and other connectors, no middleware in between.
Every Fireworks AI node, ready to drop on the canvas
16 nodes, generated from the connector itself so this list always matches what ships. Expand a category to browse.
Read and query nodes 8
Fireworks: Get Account
Get one Fireworks account: display name, account type, state, and spend-notification settings.
Fireworks: Get Dataset
Get one dataset: state, format, example count, and token estimate. Poll this after an out-of-band upload to wait for state READY before starting a fine-tuning job.
Fireworks: Get Fine-Tuning Job
Get one fine-tuning job: state, progress percent, estimated cost, and the outputModel to call once it reaches JOB_STATE_COMPLETED.
Fireworks: Get Model
Get one model's details: state, kind, context length, and capability flags such as tool support, LoRA support, and serverless availability.
Fireworks: List Accounts
List the Fireworks accounts the API key can access, with each account's ID, type, and state. Works with no parameters, so it doubles as the credential health check.
Fireworks: List Datasets
List the datasets under a Fireworks account, with each dataset's state, format, and example count. Datasets are the training inputs for fine-tuning jobs.
Fireworks: List Fine-Tuning Jobs
List the supervised fine-tuning jobs on the account with their state, base model, output model, and estimated cost.
Fireworks: List Models
List models under a Fireworks account. Use account ID 'fireworks' to browse the shared public model library, or your own account ID to list your uploaded and fine-tuned models.
Action nodes 8
Fireworks: Create Chat Completion
Run a chat completion on Fireworks serverless inference. Send a messages array, get the assistant reply back at choices[0].message.content plus token usage. Supports tool calling, JSON-mode and JSON-schema structured output, vision models via image_url content parts, and reasoning models. Non-streaming.
Fireworks: Create Completion
Run a raw text completion: prompt in, continuation out at choices[0].text. Use this for base models or when you control the full prompt format yourself; use Create Chat Completion for instruct models. Non-streaming.
Fireworks: Create Embeddings
Turn text into embedding vectors for search, clustering, or similarity. Accepts a single string or a JSON array of up to 2048 strings; vectors come back at data[].embedding in input order.
Fireworks: Create Fine-Tuning Job
Start a supervised fine-tuning (LoRA) job over an uploaded dataset. Training is billed by tokens processed, so this spends real money. The job runs asynchronously: poll Get Fine-Tuning Job for state and the resulting outputModel.
Fireworks: Create Message (Anthropic Format)
Run a Fireworks model through the Anthropic-compatible Messages endpoint: Messages-format request in, content blocks out at content[0].text. Useful when a flow or prompt library is already built around the Messages shape. Auth stays the Fireworks Bearer key. Non-streaming.
Fireworks: Delete Dataset
Permanently delete a dataset from the Fireworks account. This cannot be undone.
Fireworks: Delete Fine-Tuning Job
Permanently delete a supervised fine-tuning job record. This cannot be undone.
Fireworks: Rerank Documents
Rerank a list of documents by relevance to a query using a Fireworks reranker model. Returns each document's index and relevance_score sorted best-first: the standard second stage after a keyword or vector search.
Connecting Fireworks AI takes one credential
Fireworks authenticates with a single API key created at app.fireworks.ai. The same key covers both surfaces the connector uses, serverless inference and the account control plane, so there is no second field, no organization header, and no OAuth flow to complete.
At app.fireworks.ai, open API Keys, create a key, and copy it.
In Mattering Automations, open Connectors, choose Fireworks AI, and create a credential holding the key.
Click Connect. All 16 Fireworks nodes appear in the canvas palette.
Drop Fireworks: List Accounts on a canvas and run it as a test step. It takes no parameters and is read-only, so a successful run proves the key works and returns your account ID, which the other control-plane nodes take as account_id.
Common uses for the Fireworks AI MCP
Generation steps on open-weight models
Drafting, rewriting, summarizing, classifying, and structured extraction as one node between your data source and its destination. The chat completion node supports tool calling, JSON mode and JSON-schema structured output, vision models via image_url content parts, and reasoning models.
Search that ranks by meaning, then by relevance
Turn text into vectors with the embeddings node, up to 2048 strings per call, then use the rerank node as the second stage: hand it a query and a list of candidate documents and it returns each one's relevance_score sorted best-first, ready for a workflow to act on the top results.
A Messages-shaped drop-in for existing prompts
If your prompts or flow logic are already built around the Anthropic Messages request shape, the Create Message node runs Fireworks models through a Messages-compatible endpoint. Same request format, same content-block response, still authenticated with the one Fireworks key.
Choosing and checking models per step
Fireworks: List Models browses the shared public library with no setup, and Fireworks: Get Model reports a model's state, context length, tool support, and serverless availability, so a workflow can check a model before sending it a long prompt or a tool-calling request.
Driving the fine-tuning lifecycle
Poll a dataset until it reaches READY, start a supervised fine-tuning job over it, poll the job for progress and estimated cost, then call the resulting outputModel from the chat completion node. Training is billed by tokens processed on your Fireworks account, and the node says so up front.
Raw completions when you own the prompt format
The plain completion node takes a prompt and returns the continuation, with no chat template applied. Use it for base models, or anywhere you control the full prompt format yourself and the chat shape would only get in the way.
Example: Support questions answered from your own documents
Built on the Automations canvas in a few minutes, or described in plain language to Matter and built for you.
Trigger: a scheduled run picks up new question rows from a Database.
Fireworks: Rerank Documents scores your knowledge-base passages against the question and returns them best-first by relevance_score.
Fireworks: Create Chat Completion drafts an answer from the top passages, with JSON-schema structured output so the answer and a confidence field come back as clean fields.
The draft is written back to the row and a review task is created in Mattering Productivity, so a person approves the answer instead of writing it from scratch.
Fireworks AI MCP questions, answered
Can an AI agent run Fireworks inference through this MCP?
Yes. Matter and any MCP-connected assistant, including Claude and ChatGPT, can run chat and text completions, create embeddings, rerank documents, browse the model library, and drive the fine-tuning lifecycle. They use the same 16 nodes the automation canvas uses.
Does this work with Claude, ChatGPT, or other AI assistants?
Yes. Mattering exposes these nodes over MCP, an open protocol for connecting AI assistants to tools, so any MCP-capable assistant can connect to your workspace and reach Fireworks through them. Matter, the assistant built into the platform, uses the same nodes natively. Nothing here is tied to one assistant: the point of the connector is to give whichever assistant you already work with access to Fireworks inference.
How do I find which models I can call?
Run Fireworks: List Models. Fireworks does not expose a models listing on its inference API; the catalog lives on its control plane, organized by account. The node's account_id defaults to fireworks, the shared public library, so it returns the serverless catalog with zero configuration; set it to your own account ID, which Fireworks: List Accounts returns, to list your uploaded and fine-tuned models. Every model field on the inference nodes is free text with a documented default rather than a fixed dropdown, because the library changes on a rolling basis.
Can it generate images or transcribe audio?
No, and that reflects Fireworks rather than the connector. Fireworks deprecated its audio API in June 2026, and its current API reference no longer documents image generation, so nodes for either would be calls against endpoints that no longer answer. What Fireworks serves today is what the 16 nodes cover: text generation, embeddings, reranking, and the fine-tuning control plane. Vision still works on the input side: chat models that accept images take image_url content parts inside the messages array.
What is the Anthropic-format Messages node for?
Reuse. Fireworks documents a Messages-compatible endpoint alongside its OpenAI-compatible one, so a prompt library or flow already built around the Messages request shape can run on Fireworks models without being rewritten. Two things to know: authentication stays the Fireworks bearer key, not an Anthropic key, and the model field takes a Fireworks model ID. The reply comes back as content blocks, with the text at content[0].text.
Does the fine-tuning node spend real money?
Yes. Fireworks: Create Fine-Tuning Job starts a supervised LoRA training run that Fireworks bills by tokens processed, so treat it like the paid operation it is. The job runs asynchronously: poll Fireworks: Get Fine-Tuning Job for its state, progress percent, and estimated cost, and read outputModel once it completes. The training dataset has to be uploaded to Fireworks out of band and be in the READY state first, and the two delete nodes are permanent.
What credential does the connector use, and who bills the usage?
One Fireworks API key, sent as a bearer token and stored encrypted in Google Secret Manager. Fireworks bills the tokens and training to that key on your own Fireworks account, under the plan you hold there. That is separate from Mattering credits, which are pooled at the organization level and cover running the automation itself.
What is not covered by the 16 nodes?
A few things, each for a concrete reason. Creating and uploading datasets needs a real JSONL file upload and the canvas config panel has no file input, so uploads happen out of band with firectl or the Fireworks SDK; the list, get, and delete half of the datasets API is shipped, so the fine-tuning lifecycle runs end to end from an uploaded dataset onward. Dedicated deployments are left out because they provision billable GPU capacity, which is infrastructure administration rather than workflow automation. DPO and reinforcement fine-tuning jobs and batch inference are not in this first version. Streaming is also out of scope: every node returns one complete response, and the connector strips stream from outbound bodies so a payload pasted over from an OpenAI integration cannot silently break a node.
Put Fireworks AI to work inside your workspace
Connect once, then use every node in visual automations and AI agent workflows across the Mattering platform.
All product names, logos, and brands are property of their respective owners and are used for identification purposes only. Use of these names, logos, and brands does not imply endorsement or affiliation. Fireworks AI is a product of its respective owner; Mattering is an independent platform.