Cerebras

Run chat and text completions on Cerebras, the wafer-scale inference platform for open-weight models. Agents call the OpenAI-compatible API for high-throughput generation, use tool calling and structured outputs, and discover available models.

5 actions API key available

Cerebras website ↗ Platform Documentation ↗ Capability data verified 2026-07-25

A nightly export drops the day's support transcripts into storage for classification

List Models confirms the chosen open-weight model is still served before the run commits to it

Chat Completion processes a sample batch in JSON mode against the extraction schema

The operations lead reviews the sampled output, because at this throughput a bad prompt corrupts the whole set before anyone notices

Chat Completion runs the full set at wafer-scale speed, per-request timing recorded from time_info

The agent validates that every response parses against the schema, quarantining failures instead of writing them

Clean structured rows write back to the warehouse

A run summary with token usage, timing, and quarantine count posts to the operations channel

What This Integration Enables

Cerebras sells one thing: speed. Open-weight models, Llama, Qwen, GPT-OSS, GLM, served on wafer-scale hardware through an OpenAI-compatible API, with some of the fastest token throughput available. Speed is not a vanity metric in workflow automation; it decides which designs are possible. An agent loop that iterates five times is only viable if each iteration is fast. An inline classification step in a customer-facing flow is only viable if nobody waits for it. FlowRunner agents use Cerebras where the model call must keep up with the workflow around it. - Chat Completion covers the common case: prompt in, text out, with JSON mode, reasoning effort control, and timing data on every response - Chat Completion (Advanced) exposes the full surface: multi-turn messages, tool calling, json_schema response formats, service tiers, and sampling controls - Text Completion serves continuation and few-shot patterns where chat roles add unwanted structure - The model list populates live, so newly served models appear without a connector update - time_info on every response lets flows verify the speed advantage with their own numbers - High-volume write-backs run behind [human-in-the-loop](/concepts/human-in-the-loop) sample gates

Without FlowRunner

Inference is the slow step The classification pass over a big record set runs overnight, so the workflow around it runs weekly

Interactive flows cannot afford a model call Anything customer-facing skips the LLM step because users will not wait for it

Batch errors surface after the batch A subtly wrong prompt processes everything before anyone reads a single output

With FlowRunner

The model keeps up with the flow High-throughput generation makes the inference step cheap enough to put everywhere it helps

Real-time flows get real intelligence Classification and drafting run inline, fast enough that the user never feels the model in the loop

Speed gets a seatbelt Sample gates and schema validation stand between fast generation and the records it changes

Use Case Scenarios

Triage that finishes before the customer refreshes the page

Inbound tickets in [Freshdesk](/integrations/freshdesk) hit a Chat Completion step that classifies urgency, extracts the product area, and drafts a first response, fast enough to run synchronously in the intake flow. Urgent classifications page the on-call channel in [Slack](/integrations/slack) with the draft attached for a human to approve and send. The customer sees a fast, accurate first touch; the agent on duty sees a queue that arrives pre-sorted with the thinking already started.

A lookup loop where the model decides what it needs

A reply-drafting flow runs Chat Completion (Advanced) with a lookup tool defined. When the model requests a contact record, the flow executes Get Contact By Email against [HubSpot](/integrations/hubspot), feeds the result back, and the second pass drafts a reply grounded in the account's actual history. Because each round trip is fast, the two-call pattern stays interactive. Drafts to named customers still queue for the owner's approval before anything sends.

The backfill nobody wanted to do, done by lunch

Operations needs product categories and sentiment extracted from sixty thousand historical order notes in [BigQuery](/integrations/bigquery). The agent samples two hundred rows through Chat Completion in JSON mode, posts the sample for review, and on approval runs the remainder at full throughput, schema-validating every response and quarantining the misfits. Results write back with a confidence field, and the quarantine queue gets a human pass. The project that was quoted at three weeks of intern time closes the same day.

Human-in-Loop Highlight

The dangerous operation on this page is not one action; it is one action multiplied. Chat Completion at Cerebras throughput can process a record set faster than any human can notice the outputs are wrong, which means a subtly broken prompt, a schema that lets a wrong field through, a model that drifts on an edge case, does its damage at scale before the first spot check. So FlowRunner flows built on this connector run the sample gate: the full-volume run does not start until a person has read a statistically honest sample of what the model produced and approved it. "Two hundred sampled extractions attached, four flagged uncertain. Authorize the remaining fifty-nine thousand?" The same discipline holds for Chat Completion (Advanced) tool calls, which execute only against tools the flow explicitly defined, with consequential calls routed for approval. Speed is the product; the gate is what makes the speed safe to use on records you care about.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

5 actions

Chat

Chat Completion Generates a response for a single prompt on Cerebras wafer-scale hardware, with an optional system prompt, JSON mode for structured output, and reasoning effort control for reasoning-capable models. Returns the text, extracted reasoning where exposed, finish reason, token usage, and per-request timing.
Chat Completion (Advanced) Sends a fully custom request: a complete messages array for multi-turn conversations and content parts, tool and function calling passthrough, structured outputs via json_object or json_schema response formats, reasoning effort, service tier selection, penalties, and sampling parameters. Returns the raw API response including choices, tool calls, usage, and timing.

Completions

Text Completion Completes a raw text prompt through the OpenAI-compatible completion endpoint, without chat formatting, for continuation, fill-in, and few-shot patterns. Uses its own Max Tokens limit and a temperature range of 0 to 1.5.

Models

List Models Lists all models currently served through the Cerebras Inference API with IDs, owners, and creation timestamps. The discovery step before committing a run to a model.
Get Model Retrieves a specific model's details by ID, confirming it is still served before completion requests target it.

Frequently Asked Questions

What can FlowRunner do with Cerebras?

FlowRunner agents can run Chat Completion, Chat Completion (Advanced), and Text Completion in Cerebras, plus 2 more actions.

Does connecting Cerebras to FlowRunner require OAuth?

No. Cerebras connects to FlowRunner with an API key, no OAuth flow required.

Can Cerebras trigger a FlowRunner workflow automatically?

Cerebras doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with Cerebras

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us