Skip to content
Patchbayagent routing bay
Sign inGitHub
Menu

Agent routing bay · open source

Stop sending every task
to one expensive model.

Patchbay is a routing patchbay for agent workloads. Import live model pricing from OpenRouter and models.dev, patch models onto fast / balanced / deep lanes, and compile a deterministic routing score with an auditable SHA-384 seal chain. Your keys stay in your database, encrypted.

Patch your first agentBrowse 432 live modelsStar on GitHub
Agents in this session
0
Priced models
432
Open-weight models
18
Engine
policy-compiler@1.0.0

Live rack readout

Live· 432
  • Fast laneMistral: Mistral Nemo$0.029/MTok
  • Balanced laneMistral: Mistral Medium 3.1 (batch)$0.2/MTok
  • Deep laneOpenAI: o1-pro$150/MTok

Fetched 2026-10-11 12:57 UTC from OpenRouter and models.dev · 458 upstream records normalized. Cheapest priced model: Mistral: Mistral Nemo at $0.029/MTok in.

A real loop, not a dashboard

Every step below is a persisted operation. Nothing on this page is decorative.

1. Import real models

The catalog normalizes live OpenRouter pricing and models.dev capabilities into one shape, with attribution and a live/fallback label.

Open catalog

2. Create and patch

Create an agent, then patch models onto fast / balanced / deep lanes. Each patch writes a row, a revision, and an audit event.

New agent

3. Compile and score

Seven weighted factors produce a versioned score with itemized contributions, warnings, and a recommendation per lane.

Run the compiler

4. Export and verify

Download a routing manifest, or drive the same tools over JSON-RPC. Replay the seal chain to prove nothing was rewritten.

Export a manifest

What this is honest about

Prices are reference data

Catalog numbers come from OpenRouter and models.dev as fetched. They are not quotes, and your negotiated rate may differ.

Stored keys stay yours

Provider keys are encrypted with AES-256-GCM before they reach the database and are never returned by any read endpoint. Rotating the encryption secret invalidates them.

The score is a model, not a benchmark

policy-compiler@1.0.0 is a transparent weighted rubric over your lane assignment. It has no knowledge of your prompts, and it does not run your traffic.