nodecode-groq
Groq: fast inference for open models, each model sent only the reasoning effort it accepts. A Nodecode rewrite of AhmetArif0/hermes-groq-provider by AhmetArif0.
Groq speaks the OpenAI chat wire at https://api.groq.com/openai/v1. With
this cell, /connect offers Groq, /models lists the models that can hold a
conversation, and a turn on one of them goes out in the form Groq takes:
reasoning_effortis lowered to the nearest level the model accepts, never raised: gpt-oss takeslow,mediumandhigh, sominimal,xhighandmaxbecome the closest of those, andqwen3.8also takesnone- turning reasoning off sends
nonewhere the model has it and its weakest level otherwise (gpt-oss always reasons, so it getslow) - a model without reasoning is sent no
reasoning_effortfield at all, which Groq would otherwise refuse - whisper, orpheus, prompt-guard and allam are left out of the picker: speech models, a 4K-window model and a classifier cannot run a conversation
The model list is Groq's own /v1/models, read once at start and refreshed
hourly, with models.dev's rows for context, prices and rungs where it
publishes them. Before either has been read, the list is the three chat models
the original named.
Get it
Tell the agent "add the groq cell", or run nodecode add nc://groq.
Set it up
Make an API key at console.groq.com/keys,
then pick Groq in /models and paste the key when the picker asks, or set
GROQ_API_KEY in the environment. The address is optional, for a relay or a
proxy:
{
"groq": {
"base_url": "https://api.groq.com/openai/v1"
}
}
Use it
Ask for a model — /models lists what Groq currently serves — then talk as
usual. /think offers the rungs the chosen model takes; a pick it does not
take is lowered to one it does before the request leaves.
Groq's free plan allows 8,000 tokens a minute per model, and a turn with tools is usually larger than that, so agent work wants the paid Developer plan. That is Groq's limit; the cell does not change a request's size.
Test
From this repository's root:
.github/test-cell.sh providers/nodecode-groq
MIT licensed. See NOTICE.