A router that checks the cheap model's homework before serving it

What it does

Misceo sits between your Anthropic-compatible client (Claude Code is the obvious target) and your model backends. Eligible requests go to a lower-cost model first; a structural gate rejects empty output, upstream failures, and malformed tool calls, and an optional LLM judge scores the candidate against the user’s turn. If the answer doesn’t pass, it’s discarded and a stronger backend regenerates the served answer. Three postures — quality-first, balanced, savings-first — control how strict the judge is and what happens if it’s unavailable.

The interesting bit

Most routers decide before generation. Misceo also decides after, inspecting the completed candidate before choosing what the client receives. And it’s unusually careful about agent state: a backend that opens a tool_use/tool_result loop keeps ownership until it’s done, and cross-model handoffs handle provider-specific thinking signatures and cache markers at model-family boundaries. The first visible reply in a session is protected onto the strong backend, so your opening message doesn’t get fumbled by the cheap model.

Key highlights

• One Anthropic-compatible endpoint fronting Claude, GLM, and custom backends
• Structural validation plus optional judge, with three explicit cost/quality postures
• Session-aware handoffs, tool-loop ownership, and first-visible-reply protection
• Embedded local dashboard: live traffic, history, cost fields, judge scores, route reasons
• Manual backend handoff from the CLI or dashboard; escalated conversations stay on the strong backend for a configurable number of turns

Caveats

• No source code in the repo — it’s distributed as prebuilt npm binaries under FSL-1.1-ALv2, so review the license before redistribution or competing use
• “Local” means the proxy and dashboard run locally, not offline: inference still goes to your configured providers, and the judge sees the latest user turn, candidate, and tool names
• The dashboard has no login; keep it on loopback
• No published savings or quality numbers — the README is refreshingly honest that results vary and no benchmark results are public yet

Verdict

Worth a look if you run Claude Code (or any Anthropic-compatible agent) and want explicit, observable cost-versus-quality policy instead of a blind router. Skip it if you need open source you can audit or modify — the binary-only distribution is the deal-breaker there.