You define two destination pools: local — self-hosted, OpenAI- or Anthropic-compatible endpoints like a vLLM fleet, on infrastructure you control — and frontier — the third-party providers you already pay for. The router reads each request's shape and sends it to the cheapest pool that can actually do the job. Requests pass through Ciphero's router to get there; it keeps the decision and a digest of the prompt, never the prompt itself.
Light work stays home: short prompts, no tools — served by your own models at your own prices.
Agent work can stay home too: Theseus does the tool calling for inside models that lack it. Claude Code and Cursor get native tool calls from a stock vLLM, Ollama or llama.cpp server — no server flags, no redeploy. On Enterprise, every call is reviewed before the agent runs it.
Past the escalation point, the frontier: large context and the hardest reasoning go where they'll succeed, and the decision says why.
Degraded pools fail over: if a destination fails, new conversations go to the other pool for the next five minutes and the decision says so. Running conversations keep their model, and flagged content never leaves to fail over.
