Nautir Documentation

Documentation for Nautir, DevZero's LLM inference platform -- the AI Gateway, AI spend, routing, guardrails and evals.

Nautir Documentation

Nautir is DevZero's inference platform. It puts a gateway in front of your LLM traffic, and then uses what that gateway sees to answer four questions about your AI spend:

  • Where is the money going? Spend, savings and waste attributed to people, teams, services and departments.
  • Which model should answer? Routing profiles, fallback chains, provider preferences and auto mode.
  • What is allowed? Guardrails covering budgets, rate ceilings, model and provider access, and content rules.
  • Would a different model be better? Trials, experiments and datasets that compare candidates against your own traffic and judge the results.

Start here

If you want toRead
Understand what the platform is and how a request flows through itWhat is Nautir
Learn the product's vocabulary before anything elseKey concepts
Get traffic flowing through a gatewayGateway quickstart
Call the gateway from an SDKProduct surfaces and Request headers
Read your spend and savings honestlyAI spend
Control which model serves a requestRouting profiles
Put budgets, access rules and content filters on trafficGuardrails
Compare models on your own trafficEvals

This documentation describes behaviour established in the product's own source of truth: the gateway's published contract, the control-plane API schemas, and the platform architecture guide. Where a figure depends on your own deployment -- a hostname, a chart value, a limit you set -- the page says so rather than inventing a default.

Where the product lives

The platform has two halves and you will meet both.

The gateway is the process your LLM clients talk to. It runs either as a DevZero-hosted deployment or inside your own environment, and it is the only component that sees your prompts.

The control plane is what the dashboard talks to. It holds your guardrails, routing profiles, keys, catalog, trials and telemetry, and it never sits in the request path.

See How it works for the order a single request passes through.

On this page