Getting started.

Nexus sits between your application and the model providers. You point your client at the gateway, declare routing policy, and wire evaluation gates into CI.

Install

npm install @nexus/runtime
# or
pip install nexus-runtime

Configure routing

import { nexus } from "@nexus/runtime";

export default nexus({
  routes: [{
    when: { task: "support-reply" },
    prefer: "fast",
    maxCostPer1k: 0.9,
    fallback: ["gpt-class", "claude-class"]
  }],
  guard: {
    redact: ["pii", "secrets"],
    blockOnEvalDrop: 0.04
  },
  telemetry: { tenant: true, audit: true }
});

Evaluation gates

Each release runs against a recorded traffic sample. A drop beyond your threshold blocks the deploy:

nexus eval gate --release $RELEASE --threshold 0.04

Gateway endpoints

• POST /v1/chat/completions — OpenAI-compatible. Existing SDKs work unchanged.
• POST /v1/route — policy-based routing across providers.
• GET /v1/usage — per-tenant spend and quality metrics.

Requirements

Node 20+ or Python 3.11+. A provider API key. No additional infrastructure — the gateway is stateless and runs wherever your service runs.