Z.AI Proxy

A simple OpenAI-compatible API proxy for Z.AI, designed for deployment on Cloudflare Workers or Docker.

Features

• OpenAI-compatible API - Works with any OpenAI client library
• Token pool load balancing - Automatically manages multiple tokens for better performance
• Simple authentication - No credentials needed, tokens fetched automatically
• Streaming support - Full support for streaming chat completions
• Docker ready - Easy deployment with Docker containers

Quick Start with Docker

1. Build and Run

# Build the Docker image
docker build -t z-ai-proxy .

# Run with default settings (5 token pool)
docker run -p 3000:3000 z-ai-proxy

# Run with custom token pool size
docker run -p 3000:3000 -e TOKEN_POOL_SIZE=10 z-ai-proxy

2. Using Docker Compose

# Start the service
docker-compose up -d

# View logs
docker-compose logs -f

# Stop the service
docker-compose down

Development Setup

Local Development

# No dependencies needed - uses only Node.js built-ins

# Run tests
TOKEN_POOL_SIZE=2 node test.js

# Start development server
TOKEN_POOL_SIZE=5 node server.js

# Start with different port
PORT=8080 TOKEN_POOL_SIZE=3 node server.js

Cloudflare Workers Deployment

# Deploy the same index.js to Cloudflare Workers
# Set TOKEN_POOL_SIZE environment variable in Workers dashboard
# Deploy index.js as your worker script

Configuration

The service is configured via environment variables:

Variable	Default	Description
TOKEN_POOL_SIZE	5	Number of tokens to maintain in the pool
API_KEY	sk-z2api-key-2024	API key for client authentication
SHOW_THINK_TAGS	false	Whether to show thinking tags in responses
PORT	3000	Port to run the service on

API Usage

The proxy provides OpenAI-compatible endpoints:

Chat Completions

curl -X POST http://localhost:3000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-z2api-key-2024" \
  -d '{
    "model": "glm-4.5-air",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": false
  }'

Available Models

curl http://localhost:3000/v1/models \
  -H "Authorization: Bearer sk-z2api-key-2024"

Supported Models

• glm-4.5 - Full model
• glm-4.5-air - Lightweight model

How It Works

1. Token Pool Initialization: On startup, the service fetches N tokens (configurable) from Z.AI
2. Load Balancing: Requests are distributed across tokens using least-recently-used algorithm
3. Automatic Refresh: Failed tokens are automatically refreshed in the background
4. OpenAI Compatibility: Request/response format matches OpenAI API exactly

Architecture

Client Request → API Key Auth → Token Pool → Z.AI API → Response Processing → Client

The service maintains a pool of Z.AI tokens and automatically:

• Fetches fresh tokens when needed
• Tracks token usage and failures
• Provides load balancing across the token pool
• Handles both streaming and non-streaming responses

Testing

The service includes comprehensive tests:

# Run all tests
node test.js

# Test with different pool sizes
TOKEN_POOL_SIZE=10 node test.js
TOKEN_POOL_SIZE=1 node test.js
TOKEN_POOL_SIZE=0 node test.js  # Should fail gracefully

License

MIT