> ## Documentation Index
> Fetch the complete documentation index at: https://docs.atako.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Groq

> Connect Groq to your Atako agents — 2 read and 3 write actions.

Let your agents run fast inference on open models hosted by Groq — chat completions, the Responses API, document reranking and the model catalogue.

## Connection

* **Authentication**: API key (API key (gsk\_...)).

<Note>
  console.groq.com → sign in → pick the project in the Projects dropdown at the top (keys are bound to the selected project) → "API Keys" (console.groq.com/keys) → "Create API Key". Give it a name, then copy the key immediately (it starts with gsk\_) — it is not shown again. Groq keys have no per-endpoint permissions; the models the key can call follow your organization and project model permissions.

  See [Groq's documentation](https://console.groq.com/keys).
</Note>

## Read actions (2)

| Action | Description |
| - | - |
| `get_model` | Fetch metadata for a single model by id (e.g. "llama-3.3-70b-versatile"), as returned by list\_models. |
| `list_models` | List the models available to the API key (ids to pass as "model" to the other actions). |

## Write actions (3)

| Action | Description |
| - | - |
| `create_chat_completion` | Generate a model reply for a conversation (non-streaming, no tools). Arguments: model (string, a model id from list\_models, e.g. "llama-3.3-70b-versatile"); messages (array of at least one object \{role: "system"\|"user"\|"assistant", content: string (plain text), name?: string} — the conversation so far, oldest first, usually one optional "system" message then alternating "user"/"assistant", ending with a "user" message); temperature (number 0–2); top\_p (number 0–1); max\_completion\_tokens (integer ≥ 1); stop (a string, or an array of up to 4 strings); seed (integer); response\_format (object — \{type: "text"}, \{type: "json\_object"} (the messages must also ask for JSON), or \{type: "json\_schema", json\_schema: \{name: string, schema?: object (a JSON Schema), description?: string, strict?: boolean}} for Structured Outputs on models that support it). |
| `create_response` | Generate a model response with the OpenAI-compatible Responses API (non-streaming, no tools, nothing stored). Arguments: model (string, a model id from list\_models); input (either a plain string, treated as one user message, or an array of at least one object \{role: "system"\|"developer"\|"user"\|"assistant", content: string (plain text)}); instructions (string — a system message inserted first); max\_output\_tokens (integer ≥ 1, includes reasoning tokens); temperature (number 0–2); top\_p (number 0–1); text (object \{format: \{type: "text"} \| \{type: "json\_object"} \| \{type: "json\_schema", name: string, schema: object (a JSON Schema), description?: string, strict?: boolean}}). |
| `rerank_documents` | Rank documents by relevance to a query (scores sorted in descending order). Arguments: model (string, a reranking model id from list\_models); query (string — the search query); docs (array of 1 to 100 non-empty strings — the documents' text); instruction (string, optional — guides the ranking; a default is used otherwise). |

## Permissions

Every action above must be explicitly granted to an agent before it can be used. See [Permissions](/integrations/permissions) for the grant model and [Security](/integrations/security) for how credentials are protected.

***

*Last reviewed against the provider API: September 2026.*


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.