Universal AI API Gateway, Protocol Translator & Token Optimization Proxy
Connect any AI developer tool or client (Claude Code, Cursor, OpenCode, Codex, Cline, Kilo, etc.) to 150+ AI providers through a unified OpenAI-compatible endpoint.
Quick Start • Architecture • Features • Client Tools • Configuration
VansRouter sits between your AI coding tools and upstream model providers. It acts as an intelligent proxy that handles protocol translation, credential rotation, token compression, and multi-tier failover.
Client (OpenAI / Claude format)
│
▼
POST /v1/*
│
┌───────┴─────────────────────────────────────────────┐
│ VansRouter │
│ │
│ 1. Auth & ACL (Verify API key, check ACL) │
│ 2. Token Saver (RTK context compression) │
│ 3. Model Routing (Direct, Combo, or Failover) │
│ 4. Translation (OpenAI ↔ Claude ↔ Gemini) │
│ 5. Circuit Breaker (Proxy & Account Health) │
└───────┬─────────────────────────────────────────────┘
│
▼
Upstream Provider (Anthropic, OpenAI, Meta, Google, DeepSeek, etc.)
Requirements: Node.js >= 22.5.0
npm install -g vansrouter
vansrouterThe Web Dashboard will launch at http://localhost:20128/masuk.
(On macOS without NVM, run sudo npm install -g vansrouter --prefer-online if your global node_modules directory requires root permissions).
docker run -d \
--name vansrouter \
-p 20128:20128 \
-v 9router-data:/app/data \
-e INITIAL_PASSWORD="$(openssl rand -base64 24)" \
--restart always \
ghcr.io/vanszs/vansrouter:X.Y.ZSave the generated INITIAL_PASSWORD securely. If you omit INITIAL_PASSWORD, a new installation uses the compatibility default 123456; change it before public exposure. Open http://localhost:20128/masuk and log in.
git clone https://github.com/Vanszs/VansRouter.git
cd VansRouter
pnpm install
pnpm run build
pnpm run startExposes a single OpenAI-compatible /v1 endpoint supporting:
- Chat & Reasoning: OpenAI Chat, Anthropic Claude Messages, OpenAI Responses, and Google Gemini formats.
- Vision & Document OCR: Multimodal image-to-text understanding and document parsing (
imageToText). - Image Generation: DALL-E, FLUX, Imagen, Muse Image, and custom diffusion endpoints (
/v1/images/generations). - Speech & Audio: Real-time Text-to-Speech (
/v1/audio/speech) and Audio Transcription / STT (/v1/audio/transcriptions). - Text Embeddings: Vector embedding models across OpenAI, Qwen, Perplexity, and custom self-hosted nodes (
/v1/embeddings). - Web Search: Built-in unauthenticated local SearXNG search or provider-backed search integration.
Send standard OpenAI requests to Anthropic Claude or Google Gemini models, or vice versa, without rewriting your client code. VansRouter handles tool-call schema conversion, stream chunk demuxing, thinking tags (<think>), and token estimation automatically.
-
RTK (Real-Time Compression): Analyzes and trims repetitive tool outputs (
git diff,grep, directory trees, test logs) within context history, reducing payload sizes on long sessions. - Caveman Mode: Injects concise communication prompts to reduce verbose assistant replies.
- Ponytail Mode: Enforces lazy-senior-developer principles to minimize unnecessary boilerplate generation.
-
LoopGuard Protection: Protects against model reasoning and tool-calling loops with
$O(K)$ bounded sliding window checks, preventing event-loop stalls on 1,000+ message histories.
- Combo Routing: Combine multiple providers under a single virtual model name with customizable fallback strategies (Priority Fallback, Round-Robin, Fusion Judge, and Capacity Auto-Switch).
- Circuit Breakers: Tracks consecutive upstream failures and temporary 429 rate limits, automatically routing traffic to healthy standby accounts or proxy pools.
- Proxy Pool Integration: Bind specific providers or accounts to SOCKS5/HTTP proxy pools with health scoring.
Issue scoped API keys with granular permissions:
- Restrict keys to specific Providers (e.g. only Claude and DeepSeek).
- Restrict keys to specific Combos or Models.
- Restrict keys to specific Service Kinds (e.g., allow
llmandweb, denyimageortts).
Configure your favorite AI agent by pointing its base URL to VansRouter:
{
"provider": {
"vansrouter": {
"name": "VansRouter",
"api": "http://localhost:20128/v1"
"type": "openai",
"key": "your-vansrouter-api-key"
}
}
}export ANTHROPIC_BASE_URL="http://localhost:20128"
export ANTHROPIC_API_KEY="your-vansrouter-api-key"
claude- Open Cursor Settings → Models → OpenAI API Key.
- Set Base URL:
http://localhost:20128/v1 - Enter your VansRouter API key.
- Add desired model names (e.g.,
claude-sonnet-4.5,gpt-5.4,deepseek-v4-flash).
In ~/.codex/config.toml:
model = "claude-sonnet-4.5"
model_provider = "openai-chat-completions"
[model_providers.openai-chat-completions]
base_url = "http://localhost:20128/v1"
env_key = "VANSROUTER_API_KEY"VansRouter is configured via environment variables:
| Variable | Default | Description |
|---|---|---|
PORT |
20128 (or 3003) |
Port to listen on. |
HOSTNAME |
0.0.0.0 |
Bind host address. |
DATA_DIR |
~/.9router (macOS/Linux) or %APPDATA%\9router (Windows) |
Root directory for persistent SQLite database and backups. |
REQUIRE_API_KEY |
false |
Enforce valid Bearer API key on all /v1/* routes. |
JWT_SECRET |
Auto-generated | Secret for signing dashboard session cookies. |
INITIAL_PASSWORD |
123456 for a new installation |
Initial dashboard password; set a strong value before public exposure. |
HTTP_PROXY, HTTPS_PROXY |
"" |
Outbound proxy for upstream provider requests. |
SEARXNG_URL |
http://127.0.0.1:8888/search |
Endpoint for the local SearXNG search provider. |
GROK_CLI_VERSION |
1.0.44 |
Grok CLI version advertised to cli-chat-proxy.grok.com. The upstream answers 426 Upgrade Required below its minimum (1.0.13); raise this to clear a future gate without a code change. Read at startup — restart to apply. |
This project is licensed under the MIT License.