35 AI providers in one engine: how profClaw handles multi-provider routing

[thegdsks]

profClaw ships a smart router that picks the cheapest model for each query.

It scores message complexity on signals like: multi-step reasoning, code generation, tool orchestration, image analysis. Then routes to the right tier.

Simple question ('what is 2+2') goes to Haiku or Groq. Complex architecture question goes to Opus or GPT-4o. You don't pick manually unless you want to.

The savings add up. On a test workload of 100 messages/day:

• Without routing: ~/month (everything hits Sonnet)
• With routing: ~/month (trivial queries hit cheap models)

That is 60% savings with zero quality loss on the simple stuff.

Supported providers: Anthropic, OpenAI, Google, Ollama, Groq, DeepSeek, Azure, xAI, Mistral, Cerebras, Fireworks, Together, and 23 more.

Local models (Ollama) are always /bin/zsh. The router prefers them for trivial queries when available.

38 tests cover the routing logic. Complexity scoring, tier boundaries, provider availability, cost comparison.