NadirClaw

NadirClaw is a self-hosted LLM router that sits between a client and the model providers as a local proxy. For each request it predicts the smallest model likely to answer, sends the prompt there first, checks the response against quality heuristics (refusals, truncation, malformed JSON), and escalates to the next tier when the response falls below an acceptance threshold of 0.80. Each routed request is priced against a baseline so the saving is measured per request rather than estimated.

Classification uses a sentence-embedding centroid classifier with about 10 ms of overhead, with an opt-in fine-tuned DistilBERT three-class classifier and a bundled trained checkpoint as alternatives. Routing modifiers override the score: agentic tool loops, reasoning markers, image content, and long context force a stronger tier. A YAML rule engine adds per-prompt overrides such as forced escalation on patterns where the verifier is unreliable, and tier profiles are N-tier YAML files that hot-reload from disk.

The proxy exposes both an OpenAI-compatible chat completions endpoint and an Anthropic-compatible messages endpoint, so OpenAI-style clients (Open WebUI, Cursor, Continue, Codex, OpenClaw) and Anthropic-native clients (Claude Code) route through it with their streaming piped through unchanged. Onboarding commands write NadirClaw in as a model provider for OpenClaw, Codex, Claude Code, and Continue. Gemini models are called natively through the Google GenAI SDK, and every other model goes through LiteLLM. Ollama instances on the local network are discovered automatically, and OAuth login lets a subscription with OpenAI, Anthropic, or Google stand in for an API key.

Published measurements come from the trained classifier and DeBERTa verifier in the hosted Nadir Pro, which shares the routing architecture with the open-source core (whose classifier is the simpler binary centroid). On the RouterBench held-out set (n=11,420) the composed system reports AUROC 0.961, expected calibration error 0.016, 98.3% quality preserved against always using the top model, a 1.7% catastrophic-downgrade rate, and 60% lower cost. On RouterArena sub_10 (n=809) the composite score is 0.7118, a projected rank of #5. A contamination audit reports zero overlap between the training corpus and either held-out set.

The open-source core differs from Nadir Pro in several ways: a rule-based heuristic verifier rather than a trained cross-encoder, local JSONL and SQLite storage rather than Postgres, a terminal and local web dashboard rather than a hosted one, and no team features. Cost tracking, reports, budgets, Prometheus metrics, optional OpenTelemetry tracing, and a GitHub Action are included in the core.