Introduction

spawn is a local gateway for coding agents. It sits between your agent (Claude Code, Codex, Codex Desktop) and the model providers, classifies each task with a tiny local router, and forwards the request to the model that fits. You keep using your agent exactly as you do today; spawn just decides which model answers.

There's no new app to learn and no workflow to change. You point your agent at spawn once, and from then on every task is routed to the right model automatically.

Why spawn

Picking the model is a chore you shouldn't have to think about: Claude for design and frontend, a Codex model for engineering, something small and cheap for docs. Today that means juggling apps and switching models by hand, message to message, while you ration your best model to stay under rate limits.

spawn makes that decision for you, once per task:

• Right model, automatically. A local classifier reads the task and routes it to the right model and tier. You just prompt.
• Rate limits stop being your problem. Work spreads across tiers and providers by task, so you're not manually rationing your best model.
• Open weights, your config, your keys. The router runs locally and the policy is yours to edit. Unlike a hosted auto-router, nothing is a black box steering you toward someone else's cheapest inference.

How it fits together

spawn is a local reverse proxy that speaks the providers' own wire formats: the Anthropic Messages API and the OpenAI/Responses API. Because every serious agent lets you override its model endpoint, spawn can sit in front of them transparently:

• Classify. A tiny local router (spawn-router) reads the task at kickoff and predicts its type, complexity, and risk.
• Route. A policy maps that prediction to a concrete provider, model tier, and reasoning effort, using config you control.
• Forward. The request is forwarded upstream with your own credentials, and the response streams straight back to your agent.