LLM Config Reference

Provider chains per role, temperatures, and llm_config.py configuration.

Provider Chains per Role

Defined in core/llm_config.py. Primary provider is minimax across all roles. Fallbacks are tried in order when the primary fails or rate-limits.

PROVIDER_CHAINS = {
    "code_architect":    ["minimax", "groq/kimi-k2", "cerebras", "zhipu", "nvidia/kimi-k2.5"],
    "research_analyst":  ["minimax", "groq/llama-3.3", "cerebras", "zhipu", "nvidia/deepseek"],
    "mvp_strategist":    ["minimax", "cerebras", "groq/llama-3.3", "zhipu", "nvidia/qwen3.5"],
    "qa_reviewer":       ["minimax", "cerebras", "groq/llama-3.3", "zhipu", "nvidia/deepseek"],
    "data_engineer":     ["minimax", "cerebras", "groq/llama-3.3", "zhipu", "nvidia/deepseek"],
    "project_organizer": ["minimax", "groq/qwen3-32b", "cerebras", "zhipu", "nvidia/deepseek"],
    "narrative_content": ["minimax", "cerebras", "groq/llama-3.3", "zhipu", "nvidia/deepseek"],
    "verifier":          ["minimax", "cerebras", "groq/llama-3.3", "zhipu", "nvidia/deepseek"],
}

---

Temperatures per Role

TEMPERATURES = {
    "code_architect":    0.2,  # deterministic
    "qa_reviewer":       0.2,  # strict
    "verifier":          0.2,  # deterministic
    "data_engineer":     0.1,  # minimal variance
    "project_organizer": 0.3,  # structured
    "research_analyst":  0.5,  # balanced
    "mvp_strategist":    0.6,  # creative
    "narrative_content": 0.7,  # creative
}

---

Provider Notes

Provider	Model	Limits	Special config
MiniMax	MiniMax-M2.1	1,000 req/day free	Primary — no special config
Groq	llama-3.3-70b, kimi-k2, qwen3-32b	12K TPM	Key expires frequently
NVIDIA NIM	qwen3.5-397b, kimi-k2.5, deepseek-v3.2	1,000 NIM credits	Prefix nvidia_nim/
Cerebras	gpt-oss-120b	1M tokens/day	—
SambaNova	DeepSeek-V3.2	24K context	Backup only
Gemini	gemini-2.5-flash	20 req/day	1M context
Zhipu	glm-4.7-flash	—	extra_body={"enable_thinking": False}
OpenRouter	hermes-3-llama-3.1-405b	Free tier	—

---

TTL Backoff

When a provider fails:

TTL_BACKOFF = [5, 10, 20]  # minutes: 1st, 2nd, 3rd+ failures

After the backoff period, the provider is retried with an exponential cooldown cap. Circuit breaker state: CLOSED → OPEN → HALF_OPEN → CLOSED.

---

Using a Custom Provider Chain

from core.llm_config import LLMConfig

config = LLMConfig()
config.set_chain("code_architect", [
    "minimax",
    "groq/llama-3.3",
    "local-agi-m4max"
])

---

Environment Variables

MINIMAX_API_KEY=
GROQ_API_KEY=gsk_...
NVIDIA_API_KEY=
CEREBRAS_API_KEY=
SAMBANOVA_API_KEY=
OPENROUTER_API_KEY=
GEMINI_API_KEY=

For local nodes: LOCAL_MODEL_ENDPOINT=http://localhost:11434

---

Providers & Agents

17 agents and 11 crews

Mesh Nodes

Local node table with system prompts