The first release of Maestro: the open-source version of Sakana Fugu. A self-hostable, OpenAI- and Anthropic-compatible LLM routing brain. Route any models (open + closed) behind one endpoint, cheap-first with verify/escalate, and full per-request cost/route transparency. No GPU. MIT.
Highlights
- OpenAI (
/v1/chat/completions) + Anthropic (/v1/messages) compatible; drop into Claude Code, opencode, Cursor, the OpenAI/Vercel AI SDKs - Modes:
maestro-auto,maestro-fugu,maestro-ultra(multi-step decomposition), or pass-through any model id - Multi-provider: OpenRouter / Vercel AI Gateway / OpenAI / Groq / Together / Fireworks / DeepInfra / local (Ollama/vLLM) / custom
- Orchestration profiles (cheap/balanced/quality), dedupe cache, prompt registry, trace-viewer UI (
/ui) - Hardening: retries + fallback, timeouts, auth + rate-limit + per-key budget, trace redaction
- Transparent tool-calling pass-through (your agent keeps its tool loop); full Anthropic tool_use mapping
- Honest, reproducible offline benchmark (
npm run eval): ~92% of best-single quality at ~97% lower cost; beats random and rule-only - 73 tests; independently audited by adversarial agents
Verify it
git clone + npm install + npm run serve, then bash examples/setup-test.sh. See README, VERIFICATION.md, and ROADMAP.md.
Known limits (honest)
Early build. Streaming tool-calls are best-effort; the executable code verifier needs a real sandbox (see docs/SANDBOX.md); npm package publishing soon.
Not affiliated with Sakana AI. Built from the public TRINITY (2512.04695) and Conductor (2512.04388) papers.