For the AI behavior engineer

One address for every model your software calls

Polypus is an OpenAI-compatible gateway for developers shaping how their AI behaves. Your apps and agents call one address, and you change models, routes and keys in one place.

Run your own Read the docs GitHub

Swap the model, not the app

Moving to another model, provider or key is a gateway config change. Your apps keep calling the same address with the same OpenAI client.

Route each turn by how it is going

Named routers pick a model per turn: a stage router that reads tool progress, an LLM classifier, escalation that latches to a strong model, and fallback.

Choose a router

See every call

Traces go to Phoenix and HyperDX, and every failed call leaves a JSON dump that an agent can read to work out what went wrong.

Keys stay out of your apps

The gateway holds every provider URL and key, read from the OS keyring. Your apps carry one client key, or none on your own machine.

Yours, on your tailnet

polypus deploy runs the whole stack in Docker on your machine. Nothing publishes a host port, so only devices on your Tailscale network can call it.

Set it up

Your client already speaks it

Point any OpenAI client at your gateway and pick a model or a router by name. The default deploy defines router/turno for chat.

curl https://polypus-laptop.tail1234.ts.net/v1/chat/completions \
  -H "Authorization: Bearer $POLYPUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "router/turno",
       "messages": [{"role": "user", "content": "Hello"}]}'

What ships today

Chat and vision, streamed, with tools
Embeddings
Speech and transcription
Batches over the provider's own batch API
SystemOne structured evaluation
Model catalogue with allow lists

Arms it can reach

OpenRouter
Claude
Gemini
Cloudflare Workers AI
LM Studio
Local MLX speech

Run your own Polypus

One guide takes you from an empty machine to a locked gateway on your tailnet, and a coding agent can walk you through it.

Start the setup