ntorga/agent-starter-kit

Orchestrates multi-model AI workflows through Markdown-based personas, rules, and skills — the Maestro decomposes requests and routes tasks to specialized agents (Architect, Coder, Reviewer, Contextualizer) based on each persona's `preferredModel` declaration. Model-agnostic dispatch supports Claude, OpenAI, Qwen, and custom providers via stdio pipes, with automatic context management, long-term memory persistence, and session recovery across interruptions. Pure natural language framework — no SDKs or runtime dependencies — designed for cost optimization by delegating token-heavy work to cheaper models while preserving premium plans for high-leverage orchestration.