// the find
NadirRouter/NadirClaw
Open-source LLM router & AI cost optimizer. Routes simple prompts to cheap/local models, complex ones to premium — automatically. Drop-in OpenAI-compatible proxy for Claude Code, Codex, Cursor, OpenClaw. Saves 40-70% on AI API costs. Self-hosted, no middleman.
NadirClaw is a local proxy that sits in front of your LLM API calls and routes simple prompts to cheaper models, complex ones to premium. It speaks both OpenAI and Anthropic wire formats, so you point Claude Code, Cursor, or any compatible client at localhost:8856 and it handles the rest. The target audience is solo developers paying their own API bills who want to stop burning Sonnet credits on 'what does this function return?'.
The cascade architecture is the right design: classify first, verify the cheap answer before shipping it, escalate only on failure. That's not the naive 'route by keyword' approach that most of these tools take. The N-tier YAML config with hot-reload is genuinely useful — you can tune the escalation thresholds without restarting the proxy. The test suite is substantial for a project this size (~50 test files covering the classifier, cascade, verifier, OAuth, streaming fallback), which suggests the internals aren't held together with string. Exposing both /v1/chat/completions and /v1/messages means you don't need separate configuration for OpenAI-native vs Anthropic-native tools.
The benchmarks are self-reported against their own held-out set, and the RouterArena submission is still a PR, not a merged result — the 0.961 AUROC claim is not independently verified yet. The PolyForm Noncommercial license is a real adoption blocker: if you're using this at a company, you need to pay, which undercuts the 'cut your bills' pitch for the audience most able to pay. The savings estimate (40-70%) is based on classifying 60-70% of prompts as simple, but agentic workloads like Claude Code lean heavy on tool loops and multi-file context, which the repo itself says gets force-escalated to the complex model — the real savings on coding workloads are probably lower than the headline. The Pro tier gates the trained DeBERTa verifier behind a hosted subscription, meaning the OSS heuristic verifier is materially weaker and the AUROC numbers advertised on the tin don't apply to what you're actually running locally.