ORProxy

A proxy for the OpenRouter API that allows deeper customisation from just the model slug.

Overview

ORProxy is a lightweight Express.js server that acts as a proxy for the OpenRouter API. It allows you to provide model customisation with additional parameters using extra tags in the model name.

Disclaimer: This project is not officially licenced, endorsed, or developed by OpenRouter or their team.

Features

• Model Parameter Parsing: Extend model names with special parameters using the $ delimiter
• Quantization Support: Specify quantization levels (int4, fp8, etc.)
• Reasoning Options: Enable and configure reasoning capabilities
• Provider Filtering: Restrict requests to specific providers
• Streaming Support: Maintains streaming capabilities for real-time responses
• Usage Tracking: Automatically includes usage information in responses
• Anthropic Cache Breakpoints: Enable context caching for Anthropic models

Usage

Install dependencies

npm install
# or 
bun install

Run the server:

node or_proxy.js
# or
bun or_proxy.js

Then, just replace all API calls from https://openrouter.ai/api/v1 to http://localhost:3001/v1. (Note the lack of /api)

By default, the server listens on port 3001. You can change this by setting the PORT environment variable:

PORT=8080 node or_proxy.js
# or 
PORT=8080 bun or_proxy.js

Docker and Docker Compose is also supported.

Parameters

The proxy extends model names with additional parameters using the $ delimiter:

{model_name}${param1},{param2},{param3}

Quantization Parameters

• int4, int8, fp4, fp6, fp8, fp16, bf16, fp32

Example: moonshotai/kimi-k2$fp8

Reasoning Options

• think - Enable reasoning
• think.1000 - Enable reasoning with max 1000 tokens (Anthropic style)
• think.high - Enable reasoning with high effort (OpenAI style)
• think.off or think.no - Disable reasoning on models that reason by default

Example: openai/o3-mini$think.high

ZDR (Zero Data Retention)

• zdr: Force request to be handled by ZDR provider

Example: moonshotai/kimi-k2-0905$zdr

Cache Mode

Adds cache breakpoints to request (useful for Anthropic models)

• cache: Use default cache mode (5 minutes)
• cache1h: Use extended cache mode (1 hour)

Example: anthropic/claude-sonnet-4.5$cache1h

Provider Filtering

Any parameter that doesn't match quantization or reasoning options is treated as a provider slug. Provider slug are obtainable by clicking the clipboard icon next to the provider name on OpenRouter.

Example: deepseek/deepseek-r1-0528$google-vertex

Combinations

These can also be combined, for example: deepseek/deepseek-chat-v3.1$think,fireworks for DeepSeek V3.1 thinking using Fireworks

Security

This proxy does not store or log any API keys or request content. All authentication headers are passed through to OpenRouter directly. All traffic is forwarded directly to OpenRouter without any telemetry.