Local Llm Router

The big idea here is Serena MCP integration, which you call first to get actual semantic code understanding (symbols, references, type info) before hitting the LLM. This cuts down on token waste and makes edits more precise since you're working with symbol-level operations instead of grep-style text matching. The implementation includes service discovery, task classification, and automatic model selection based on query type. Honestly most useful in secure or defense environments where internet-connected AI is off the table.