Who builds this, and why.

Why it exists

Agents talk to a lot of models. Every provider has its own key, its own base URL, its own idea of what a request looks like, and its own way of failing. Wiring that into each application means the routing logic lives in the application, and changing a model means a deploy.

polyrouter puts one endpoint in front of all of it. Explicit routing first, so what you asked for is what you get; fallbacks when a provider degrades; spend limits that block rather than warn; and honest cost tracking that records metadata and never your prompts.

What this project will not do

• No markup. You bring your own provider keys and pay your providers directly. Self-hosting is free and stays free.
• No prompt storage by default. Metadata only — tokens, cost, latency, and which model served the request. Body capture is self-host only, off by default, and behind an explicit consent gate.
• No telemetry home. A self-hosted instance does not phone anywhere. Your traffic goes from your box to your providers.
• No lock-in. AGPL-3.0. Run it, fork it, read every line of the routing logic.

Get in touch

Bug reports, feature requests and questions all belong on GitHub, where the answers stay public and searchable for whoever hits the same thing next.

Polyrouter Cloud

The same router, fully managed, for people who would rather not run the container. It is on the way.