A local inference adapter that routes routine OpenClaw tasks to on-device models and only calls APIs for complex ones

Running everything through cloud APIs costs money and leaks data. Local models like Gemma 4 on RTX and Zhipu's Pony-Alpha-2 handle routine agent tasks fine, but OpenClaw has no smart routing between local and remote. This adapter classifies each agent request by complexity, routes simple ones to local inference (Ollama, LM Studio, vLLM), and only escalates to Claude/GPT for tasks that need frontier capability. A 14B local model handles 80% of calls in practice, cutting costs 60-80% on typical workloads with zero data leaving the machine for routine operations.