Groq Provider Setup: Chat Models and Whisper Transcription

Groq runs open-weight models (Llama, Gemma, Kimi, Qwen, GPT OSS, and others) on custom LPU hardware, which makes inference extremely fast. Two providers come from the Groq plugin: an OpenAI-compatible chat provider and an audio media-understanding provider.

Install plugin

Get the official plugin installed:

openclaw plugins install @openclaw/groq-provider

If a Gateway is already running, the install is picked up by it automatically; if not, it takes effect the next time it starts. See Apply changes and inspect.

Getting started

Get an API key

An API key can be generated at console.groq.com/keys.

Set the API key

export GROQ_API_KEY=gsk_...

Set a default model

{
  agents: {
    defaults: {
      model: { primary: "groq/openai/gpt-oss-120b" },
    },
  },
}

Verify the catalog is reachable

openclaw models list --provider groq

Config file example

{
  env: { vars: { GROQ_API_KEY: "gsk_..." } },
  agents: {
    defaults: {
      model: { primary: "groq/openai/gpt-oss-120b" },
    },
  },
}

Built-in catalog

A manifest-backed Groq catalog of reasoning models is bundled with OpenClaw. To view the static rows that match your installed version, run openclaw models list --provider groq; for Groq's authoritative list, consult console.groq.com/docs/models.

Tip

Every OpenClaw release brings changes to the catalog. The rows your installed version knows about are listed by openclaw models list --provider groq; check console.groq.com/docs/models as well for models that were added or deprecated.

Reasoning models

On Groq's GPT OSS reasoning models, OpenClaw's shared /think levels are translated into reasoning_effort values of low, medium, or high. When /think off or /think none is set, reasoning_effort is left out of those requests instead of a disabled value being sent.

Qwen 3.6 27B on Groq handles text only and offers two wire reasoning modes. /think off or /think none results in reasoning_effort: "none" being sent; any enabled level results in "default" being sent. For Groq Qwen models with reasoning turned on, OpenClaw asks for reasoning_format: "parsed", which keeps reasoning inside thinking blocks rather than in the visible answer text. This format hook does not touch disabled reasoning or GPT OSS requests.

The shared /think levels, and how OpenClaw maps them for each provider, are described in Thinking modes.

Audio transcription

An audio media-understanding provider is also registered by the Groq plugin, which lets voice messages be transcribed through the shared tools.media.audio surface.

To use Groq as the default audio backend:

{
  tools: {
    media: {
      models: [{ provider: "groq", capabilities: ["audio"] }],
    },
  },
}

Environment availability for the daemon

When the Gateway is run as a managed service (launchd, systemd, Docker), GROQ_API_KEY has to be visible to that process, not only to your interactive shell.

Warning

Exporting a key in an interactive shell alone does nothing for a launchd or systemd daemon unless that environment is imported there as well. Put the key in ~/.openclaw/.env or set it through env.shellEnv so the gateway process can read it.

Custom Groq model ids

At runtime, OpenClaw takes any Groq model id. Prefix the exact id that Groq shows with groq/. Common cases are covered by the static catalog; ids that are not catalogued fall back to the default OpenAI-compatible template.

{
  agents: {
    defaults: {
      model: { primary: "groq/<your-model-id>" },
    },
  },
}

Related

• Model providers: how providers are picked, how model refs work, and what happens on failover.

• Thinking modes: the reasoning effort levels and how they interact with provider policy.

• Configuration reference: the complete config schema, provider and audio settings included.

• Groq Console: the Groq dashboard, its API docs, and pricing information.