LLM Backends

GuardClaw supports multiple LLM backends for risk scoring. All cloud backends use the OpenAI-compatible API format.

Quick Comparison

Backend	Privacy	Speed	Cost	Best for
Built-in MLX	100% local	Fast (Apple Silicon)	Free	Mac users, easiest setup
LM Studio	100% local	Fast	Free	Advanced local model control
Ollama	100% local	Fast	Free	Linux/Docker environments
OpenRouter	Cloud	Very fast	Pay-per-use	Best model variety
Anthropic Claude	Cloud	Fast	Pay-per-use	Highest accuracy
MiniMax	Cloud	Fast	Pay-per-use	Cost-effective cloud option
fallback	Local	Instant	Free	No LLM available, rule-only

Selecting a Backend

guardclaw config llm

The interactive picker shows available options, lets you enter API keys, and tests the connection before saving.

Fallback Mode

If no LLM is available, GuardClaw uses deterministic rule-based scoring:

guardclaw config set SAFEGUARD_BACKEND fallback

Fallback mode:

• ✅ Instant (< 1ms per call)
• ✅ Zero memory usage
• ✅ Catches known-dangerous patterns reliably
• ❌ No context understanding
• ❌ Pattern-matching only — may miss novel attacks

Use fallback as a last resort or during LLM downtime.