Multi-Model Routing

SalmAlm supports 6 AI providers with automatic task-based routing — the right model for each message.

Supported Providers

Provider	Models	Key
Anthropic	Claude Haiku, Sonnet, Opus	ANTHROPIC_API_KEY
OpenAI	GPT-4o, GPT-4o-mini, o1	OPENAI_API_KEY
Google	Gemini 2.5 Flash/Pro	GOOGLE_API_KEY
xAI	Grok	XAI_API_KEY
DeepSeek	DeepSeek Chat/Reasoner	DEEPSEEK_API_KEY
Local LLM	Ollama, LM Studio	LOCAL_LLM_URL

Auto Routing

When model is set to auto, SalmAlm classifies each message into complexity tiers:

Classification

Tier	Examples	Default Model
Simple	Greetings, yes/no, short questions	Claude Haiku (12x cheaper)
Moderate	Explanations, summaries, analysis	Claude Sonnet
Complex	Multi-step code, architecture, debugging	Claude Sonnet

Classification uses keyword matching + heuristics (code blocks, question complexity, message length).

Cost Impact

Before (always Sonnet): ~$7.09/day
After (auto routing):   ~$1.23/day  → 83% savings

Fixed Model

Set a specific model via:

• Web UI: Model panel dropdown
• Command: /model anthropic/claude-sonnet-4-20250514
• API: POST /api/model/switch

Per-Session Override

Each session can have its own model:

/model sonnet    → This session uses Sonnet
/model auto      → Back to auto routing

Provider Key Fallback

SalmAlm tries providers in order:

1. Vault-stored API key
2. Environment variable
3. CLI OAuth token (if SALMALM_CLI_OAUTH=1)

If a provider key is missing, that provider's models are excluded from routing candidates.

Configuration

Web UI

Settings → Engine Optimization → Auto Routing panel shows:

• Current tier assignments
• Per-tier model selection
• Cost comparison with OpenClaw
• Classification guide

Environment Variables

SALMALM_DEFAULT_MODEL=anthropic/claude-sonnet-4-20250514
SALMALM_SIMPLE_MODEL=anthropic/claude-haiku-3.5-20241022
SALMALM_COMPLEX_MODEL=anthropic/claude-sonnet-4-20250514