How it fits together

The orchestrator is the routing core: it assembles context, routes each task to the optimal model, and validates outputs across frontier providers (Gemini, Claude, GPT-4.1) and self-hostable open weights (Gemma, Qwen, Kimi, DeepSeek, Llama). Every tool call passes through AgentGuard for PII, secret, and prompt-injection enforcement, and the whole mesh runs over a Tailscale zero-trust network with mTLS on every hop.

Sensitive and classified workloads stay on open-weight models run on-prem or air-gapped, so no data leaves the boundary — while frontier models handle the long-context and high-accuracy work where that trade-off makes sense.