@deepseek-router/core

v1.0.12

Published

24 days ago

Production-grade multi-provider proxy for Claude Code — DeepSeek, Kimi, Qwen, Zhipu with circuit breaker and automatic failover

0High
0Medium
0Low

cgagm

deepseek claude ai-router failover circuit-breaker anthropic openai proxy llm typescript

DeepSeek Router

⚡ Quick Start

一键安装（推荐）

摸鱼 (Moyu) — $9 开箱即用

unzip cc-china-stack-*.zip && cd cc-china-stack
bash setup.sh
export DEEPSEEK_API_KEY="sk-your-key"
claude

不用写配置、不用配代理、不用理解什么是 failover。一个脚本搞定。

手动安装（免费）

git clone https://github.com/Cgagm/deepseek-router.git
cd deepseek-router
pnpm install && pnpm run build
cp router.config.example.json router.config.json
export ANTHROPIC_BASE_URL="http://localhost:8788/v1/messages"
claude

Done. Claude Code now routes through 5 providers. If one fails, the next takes over. You won't even notice.

💡 Why DeepSeek Router?

The Problem

Claude Code is the best AI coding tool on the planet. But it talks to one API endpoint. When that endpoint goes down — and it will — your conversation dies mid-sentence. Context lost. Work gone. Rage ensues.

The Solution

DeepSeek Router sits between Claude Code and 5 independent AI providers. When one fails, another takes over automatically. Your request never drops. Your conversation never dies.

Without DeepSeek Router

Claude Code  →  Provider A
                      ↓
                  500 ERROR
                      ↓
              💀 Dead. Context lost.

With DeepSeek Router

Claude Code  →  Router  →  Provider A (fail)
                      ↓
                   Provider B (fail)
                      ↓
                   Provider C ✓
                      ↓
              🎯 Response returned.

What You Get

| Feature | Without Router | With Router | |---------|:-------------:|:-----------:| | Uptime | ~99.5% (single provider) | ~99.99% (5 providers) | | Failover | ❌ Manual | ✅ Automatic, <1ms | | Provider choice | 1 | 5 (DeepSeek, Tencent, Zhipu, Alibaba, ByteDance) | | Cost control | Fixed pricing | Shop across providers | | Circuit breaking | ❌ | ✅ Prevents cascading failures | | Health monitoring | ❌ | ✅ Real-time dashboard | | Hot reload | ❌ | ✅ No restart on config change | | API format | Provider-specific | Anthropic Messages API (works with Claude Code) |

🌐 Providers

All major Chinese AI providers. Mix and match. Pay only for what you use.

| Provider | Model | ~Price (1M tokens) | Format | |----------|-------|:------------------:|--------| | DeepSeek | deepseek-chat | $0.14 | Native Anthropic | | Tencent Hunyuan | hunyuan-lite | $0.14 | OpenAI | | Zhipu GLM | glm-4-flash | $0.14 | OpenAI | | Alibaba Bailian | qwen-turbo | $0.11 | OpenAI | | ByteDance Volcengine | doubao-lite | $0.11 | OpenAI |

~$0.11–$0.14 per MILLION tokens. Claude Code's typical daily usage costs pennies.

Setup guides for each provider →

📊 Benchmarks

Scenario: Provider A rate-limits during a 10-message coding session

Without Router:
  Message 1-5: ✓
  Message 6:   429 → 💀 Session dead. Start over.

With DeepSeek Router:
  Message 1-5:  Provider A ✓
  Message 6:    Provider A 429 → Provider B ✓
  Message 7-10: Provider B ✓
  Result:       Session complete. Nothing lost.

Overhead: <1ms routing decision. Zero impact on response latency.

🏗 Architecture

POST /v1/messages
       │
       ▼
┌─────────────────┐
│   Rate Limiter  │  ← per-IP token bucket (60 req/min)
└────────┬────────┘
         ▼
┌─────────────────┐
│  Zod Validation │  ← strict request schema, 400 on invalid
└────────┬────────┘
         ▼
┌─────────────────┐
│ Failover Router │  ← priority chain, skips open circuits
└────────┬────────┘
         ▼
┌─────────────────┐
│ Circuit Breaker │  ← Closed → Open → HalfOpen → Closed
└────────┬────────┘
         ▼
┌─────────────────┐
│ Format Adapter  │  ← Anthropic ↔ OpenAI translation
└────────┬────────┘
         ▼
┌─────────────────┐
│    Provider     │  ← HTTP request with AbortSignal timeout
└─────────────────┘

📡 API

`POST /v1/messages`

Anthropic Messages API — drop-in replacement for api.anthropic.com.

curl http://localhost:8788/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: any-value" \
  -d '{
    "model": "deepseek-v4-flash",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Supports: system, stream (SSE), tools, tool_choice, stop_sequences, temperature, top_p, top_k.

`GET /health` — JSON health report

{
  "status": "healthy",
  "uptime": 3600,
  "version": "1.0.0",
  "providers": [...],
  "summary": { "total": 5, "healthy": 5, "degraded": 0, "down": 0 }
}

`GET /metrics` — Prometheus format

Compatible with Grafana, Datadog, VictoriaMetrics.

`GET /` — Human-readable status page

🔧 Configuration

Edit router.config.json. Changes take effect within seconds — no restart needed.

{
  "providers": [
    {
      "name": "deepseek",
      "endpoint": "https://api.deepseek.com/anthropic/messages",
      "apiKey": "$DEEPSEEK_API_KEY",    // env var interpolation
      "format": "anthropic",
      "authType": "x-api-key"
    }
  ],
  "router": {
    "providerOrder": ["deepseek", "tencent", "zhipu", "aliyun", "volcengine"],
    "port": 8788,
    "apiKey": "",                       // set to protect your router
    "circuitBreaker": {
      "failureThreshold": 5,
      "resetTimeoutMs": 30000
    }
  }
}

Full example with comments: router.config.example.json

🔒 Security

Auth: Optional API key protection on all endpoints (Authorization: Bearer or x-api-key)
Error sanitization: Provider API keys never leak to clients — errors return HTTP 502, not raw upstream bodies
HTTPS-only: Config enforces HTTPS endpoints. HTTP not accepted.
Header injection prevention: Custom headers validated against CRLF and reserved names
Zod validation: Every request body is strictly validated before reaching providers
Rate limiting: Per-IP token bucket (default: 60 req/min, 10 concurrent)

🧪 Quality

TypeScript strict mode  ·  158 tests  ·  CI on Node 18/22/24  ·  Vitest  ·  Docker

📦 Install Options

Option 1: Docker (recommended)

docker run -d -p 8788:8788 \
  -v ./router.config.json:/app/router.config.json \
  ghcr.io/Cgagm/deepseek-router:v1.0

Docker Hub →

Option 2: npm

npm install @deepseek-router/core

Option 3: Git

git clone https://github.com/Cgagm/deepseek-router.git
cd deepseek-router && pnpm install && pnpm run build

❓ FAQ

Q: Does this replace Claude API? No. It's a proxy. You still need API keys from the AI providers. The router makes them speak Anthropic's format and handles failover.

Q: How many providers do I need? At least 2 for failover. All 5 for maximum reliability.

Q: What's the latency overhead? <1ms. The router only makes routing decisions. Response data is passthrough.

Q: Can I use this commercially? Yes. MIT license. Build it into your product, sell it as a service — whatever you want.

Q: Why not use the official Claude API directly? You can! But if Anthropic has an outage (it happens), your work stops. This adds 4 backup paths for pennies a day.

Q: How is this different from claude-code-router? claude-code-router is great (34K+ stars), but it has no circuit breaking or automatic failover. If a provider fails, your request fails. DeepSeek Router was built for one thing: your request always succeeds.

⭐ Star History

If this project saves you from losing a single coding session, consider giving it a star.