Open-source, self-hosted LLM gateway. Point your agents at one URL with one key — explicit-first routing with fallbacks, hard spend limits, and honest, metadata-only cost tracking. Your keys, your box, zero markup.

Hi all,

I got sick of only providing a single model to Openclaw/Hermes Agent/Opencode/other harnesses without any sort of intellegent routing paradigms.

Heartbeats and simple cron jobs shouldn't use the same expensive models as complex coding or research requests. Thus, I developed "polyrouter."

The idea is simple: add tiers, fallbacks, etc, and this app will intellilgently route each request to models as you define, based on complexity - fingerprinting and removing harness system prompts (etc) so they don't get accidentally considered as "complex" - only focusing on the actual pure prompt request itself.

Full disclosure: I did take inspiration from Manifest (https://manifest.build), however they are deprecating one of their main functionalities for this purpose - rule-based routing - so I developed this app instead, while simultaneously improving the feature.

Self-host as you see fit (Docker recommended) and please provide any constructive feedback/feature requests/improvement requests - would love any and all suggestions for improvements!

finally set this up on a homelab box this weekend and the explicit-first routing feels genuinely thoughtful, you actually know which provider handled which call. spend caps saved me from a runaway agent already, huge plus for self-hosted llm work.

Finally got around to setting this up on a small box and the spend limits actually save me from surprise bills, which is rarer than it should be. Routing rules being explicit-first means I finally understand what my agents are doing at 2am.

polyrouter

One endpoint for every model.

One endpoint for every model.