Local proxy. Zero dependencies. Single binary.

Route any model
through any provider

Stop burning expensive tokens on simple tasks. Refract sits between your Anthropic-speaking tools and every LLM provider — translating, routing, and hot-swapping models mid-conversation. OpenAI incoming support coming in v1.x.

Buy — $19 See how it works

$ refract init
Created ~/.refract/config.json

$ refract start
refract service started.

$ refract claude
Claude Code connected via OpenRouter

$ /model deepseek-r1
Routed to deepseek/deepseek-r1 via openrouter

$ /model gemini-2.5-pro
Routed to google/gemini-2.5-pro via openrouter

The problem

Locked in

Your coding tool speaks one API format. Your models speak another. You pick one provider and stick with it — or write your own adapter.

Burning tokens

Not every request needs a frontier model. But there's no easy way to route simple tasks to cheap models and complex ones to capable models.

Rate limited

One provider, one rate limit. When it hits, you're stuck. No fallback, no redistribution. Just wait.

How it works

Refract is a local reverse proxy. It sits between your AI coding tools and your LLM providers, handling translation and routing so nothing else has to change.

Your tool sends a request

Claude Code, Continue — any tool that speaks the Anthropic API sends a request as usual.

Refract translates and routes

Translates Anthropic format to OpenAI format. Evaluates your routing rules. Picks the right model and provider for that specific request.

Provider responds

OpenRouter, Anthropic, OpenAI, Ollama — the response streams back and Refract translates it into Anthropic format your tool expects.

Intelligent routing

Define rules that match the right model to the right request. First match wins. No when clause = catch-all default.

~/.refract/config.json

{
  "routes": [
    {
      "name": "big-context",
      "when": { "token_estimate": { "gt": 60000 } },
      "model": "google/gemini-2.5-pro"
    },
    {
      "name": "deep-thinking",
      "when": { "thinking": { "enabled": true } },
      "model": "deepseek/deepseek-r1"
    },
    {
      "name": "default",
      "model": "google/gemini-2.0-flash-lite-001"
    }
  ]
}

token_estimate Route big context to capable models

thinking Route extended thinking to reasoning models

system_prompt Match on prompt content

model Route by requested model name

tools Match when tools are present

exec / http Custom shell or service matchers

Every provider, one config

Declare your providers once. Route to any of them by name. Mix cloud and local in the same setup.

OpenRouter openai Hundreds of models

Anthropic openai Direct API access

OpenAI openai Direct API access

Ollama ollama Local models, no key needed

More providers coming soon.

Built for how you actually work

Hot-swap mid-session

Switch models without restarting. Your tool's /model command routes through your rules on every request. Gemini to DeepSeek to Llama — change whenever you want.

Full streaming translation

Anthropic SSE in, OpenAI SSE out. Streaming, non-streaming, thinking blocks, images, tool use. All translated faithfully. No data loss.

Zero dependencies

Go stdlib only. No Node.js, no Python, no Docker requirement. Download a binary and run refract init to get started. That's the entire install process.

Built-in dashboard

See every request that passes through. Which rule matched, which provider handled it, how long it took. Live feed, route table, hit counts.

Runs as a service

refract start installs it as a system service. Runs in the background, restarts on failure, starts at login. One command to manage it all.

Env var interpolation

Reference $API_KEY in your config. No secrets in JSON files. Works with .env files, Vault, whatever you already use.

Pricing

Refract

$19

one-time payment. No subscription.

Full API translation — Anthropic to OpenAI format and back
7 routing matchers including exec and http
Multi-provider support (OpenRouter, Anthropic, OpenAI, Ollama)
Hot-swapping mid-session
CLI with service management — refract init, start, stop, claude
Built-in dashboard
Pre-built binaries for macOS, Linux, Windows
Free updates through v1.x
OpenAI incoming request support (coming in v1.x)

Buy Refract

Early access pricing. Will increase as features land.

Route any model through any provider