Groq with Empryo

Groq runs open-weight models on its custom LPU hardware at speeds no GPU host matches, and Empryo ships it as a built-in provider. Point Empryo at Groq for the parts of the agent loop where latency dominates: search, summaries, cleanup, and routing. Set one key and the whole model lineup is available.

Setup

Groq is one of Empryo's 21 built-in providers, so there is no custom config block to write. Get a key at console.groq.com (keys start with gsk_), then save it:

empryo --set-key groq gsk_...

The key goes to your OS keychain when one is available, with a 0600 file fallback otherwise. Confirm it landed:

empryo --list-providers
empryo --list-models groq

Pick a Groq model interactively by launching Empryo and pressing Ctrl+L, or set one as the default in ~/.empryo/config.json:

{
  "defaultModel": "groq/openai/gpt-oss-120b"
}

Run a one-off from the terminal with --model:

empryo --headless --model groq/llama-3.3-70b-versatile "summarize the auth module"

Where Groq shines: the task router

Empryo lets you assign a different model to each slot of the agent loop. Groq's speed is most valuable on the high-frequency background slots, so you can keep a frontier model for code writing and hand the fast, cheap work to Groq:

{
  "defaultModel": "anthropic/claude-sonnet-4-5",
  "taskRouter": {
    "spark":     "groq/llama-3.3-70b-versatile",
    "webSearch": "groq/llama-3.1-8b-instant",
    "compact":   "groq/openai/gpt-oss-20b",
    "semantic":  "groq/llama-3.1-8b-instant",
    "verify":    "groq/openai/gpt-oss-120b"
  }
}

spark runs read-only research agents, webSearch powers the search agent, compact handles context compaction, semantic writes the Genome's one-line summaries, and verify runs the adversarial review pass. These fire constantly, and Groq's throughput keeps them from being the bottleneck. Tune the rest with /router inside the TUI.

Per-launch key

For CI or one-off sessions, pass the key as an env var instead of storing it:

GROQ_API_KEY=gsk_... empryo

Nothing is persisted. If a shell-exported key keeps shadowing a stored one, set "keyPriority": "app" in config to prefer the keychain.

When to use it

Groq's pitch is one thing done extremely well: inference latency. Its LPU hardware serves open-weight models at hundreds to over a thousand tokens per second, which is several times faster than typical GPU-hosted endpoints. That makes it a strong fit for Empryo's task router, where research, web search, compaction, and Genome summary slots run on every turn and you want them to finish before they slow the loop. The current production lineup leans on OpenAI's gpt-oss-120b and gpt-oss-20b plus Llama 3.3 70B and 3.1 8B, with Qwen3 and others in preview, and the groq/compound systems add built-in web search and code execution. Be honest about the tradeoff: Groq does not host the frontier closed models (Claude, GPT-5-class, Gemini), so for the heavy code-writing slots you will usually pair it with Anthropic, OpenAI, or Google and reserve Groq for the fast lane. Preview models can be discontinued, so pin production work to the production-tier IDs. If you want speed on cheap, repetitive agent work without giving up a strong model for the actual edits, Groq is the provider to slot into the router.