routeur.ai / Documentation
Dashboard

Introduction

routeur.ai is a single OpenAI-compatible endpoint that sits between your application and the upstream LLM providers you use today. Point your existing SDK at https://api.routeur.ai/v1, send the same chat completion request you send to OpenAI, and routeur.ai picks the cheapest capable model that satisfies your routing rules, applies DLP and output moderation, and returns the standard response shape your code already understands.

What routeur.ai does on every request

In this order, on the hot path:

  1. Authenticate the caller against a routeur.ai API key.
  2. Check billing and allowance — a cancelled subscription or an exhausted allowance is refused here.
  3. Apply input DLP — redact or block sensitive substrings before anything else looks at the prompt.
  4. Apply prompt shields — detect instruction-override, encoded-payload and other injection attempts.
  5. Apply input moderation — managed categories or your own patterns.
  6. Score and gate — the auto-router scores the prompt; classifier rules can block on a threshold.
  7. Resolve a route — pick the upstream provider and model from your rules, or let auto choose.
  8. Forward and moderate — call the upstream provider, run output moderation on the result, return the standard OpenAI-shape response.

Note that the safety layers run before the routing decision, not after it.

Why a gateway

Three problems that every team running LLM features hits, and the one place to solve them.

Cost control

Centrally pick the cheapest model that meets the request — without redeploying client code.

Data safety

DLP rules redact secrets, PII, and prompt-injection vectors before the upstream model ever sees them.

Observability

Every request is traced with provider, model, latency, tokens, cost, and an optional full-payload archive.

Design principles

  • OpenAI-compatible at the edge. Existing SDKs, fewer migrations.
  • Stateless hot path. The gateway holds no per-request state. Configuration is read from a managed read model behind a short cache, so a rule change in the dashboard takes effect within seconds — no redeploy, no restart.
  • Short, JSON error bodies. routeur.ai never relays raw upstream error pages back to your application.
  • Auditable. Trace records and optional payload archives let you reconstruct any request.

Next

Quickstart →