openrouter-skill

Enable intelligent routing to Chinese/open-source LLMs through OpenRouter's unified API. Provides model selection guidance, cost optimization, and production patterns for LangChain/LangGraph integration.

<quick_start>

1. Basic LangChain Setup

from langchain_openai import ChatOpenAI
import os

# Any OpenRouter model works with ChatOpenAI
llm = ChatOpenAI(
    model="deepseek/deepseek-chat",
    openai_api_key=os.getenv("OPENROUTER_API_KEY"),
    openai_api_base="https://openrouter.ai/api/v1",
    default_headers={
        "HTTP-Referer": "https://your-app.com",  # Optional but recommended
        "X-Title": "Your App Name"
    }
)

response = llm.invoke("Explain quantum computing in simple terms")

2. Vision Analysis (Charts, Documents)

from langchain_core.messages import HumanMessage
import base64

llm = ChatOpenAI(
    model="qwen/qwen-2-vl-72b-instruct",
    openai_api_key=os.getenv("OPENROUTER_API_KEY"),
    openai_api_base="https://openrouter.ai/api/v1"
)

# From URL
response = llm.invoke([
    HumanMessage(content=[
        {"type": "text", "text": "Analyze this chart and identify key trends"},
        {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}}
    ])
])

# From base64
with open("chart.png", "rb") as f:
    image_data = base64.b64encode(f.read()).decode()

response = llm.invoke([
    HumanMessage(content=[
        {"type": "text", "text": "What does this chart show?"},
        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_data}"}}
    ])
])

3. Auto-Routing (Let OpenRouter Choose)

# OpenRouter's Auto model selects the best model for your prompt
llm = ChatOpenAI(
    model="openrouter/auto",  # Powered by NotDiamond
    openai_api_key=os.getenv("OPENROUTER_API_KEY"),
    openai_api_base="https://openrouter.ai/api/v1"
)

</quick_start>

<success_criteria>

• OpenRouter API key configured and authenticated
• Model selection follows the decision tree (vision -> Qwen-VL, code -> DeepSeek Coder, etc.)
• LangChain ChatOpenAI integration working with correct base URL and headers
• Cost savings of 60-97% vs Western model equivalents for comparable quality
• Fallback chain configured for production reliability </success_criteria>

<core_concepts>

Model Selection Decision Tree

Task Type
│
├─ Vision/Charts ──────────> qwen/qwen-2-vl-72b-instruct ($0.40/M)
├─ Code Generation ────────> deepseek/deepseek-coder ($0.14/$0.28)
├─ Deep Reasoning ─────────> qwen/qwq-32b ($0.15/$0.40)
├─ Long Documents ─────────> moonshot/moonshot-v1-128k ($0.55/M)
├─ Fast/Cheap Tasks ───────> qwen/qwen-2.5-7b-instruct ($0.09/M)
├─ General Analysis ───────> deepseek/deepseek-chat ($0.27/$1.10)
└─ Unknown/Auto ───────────> openrouter/auto

Top Chinese LLMs

Model	Best For	Cost ($/1M tokens)
deepseek/deepseek-chat	General reasoning, analysis	$0.27 in / $1.10 out
deepseek/deepseek-coder	Code generation	$0.14 / $0.28
qwen/qwen-2-vl-72b-instruct	Vision, charts	$0.40 / $0.40
qwen/qwen-2.5-7b-instruct	Fast, cheap tasks	$0.09 / $0.09
qwen/qwq-32b	Deep reasoning	$0.15 / $0.40
moonshot/moonshot-v1-128k	Long context (128K)	$0.55 / $0.55

Cost Comparison vs Western Models

Task	Western Model	Cost	Chinese Model	Cost	Savings
Chat	GPT-4o	$5.00/$15.00	DeepSeek Chat	$0.27/$1.10	95%
Code	Claude Sonnet	$3.00/$15.00	DeepSeek Coder	$0.14/$0.28	95%
Vision	GPT-4o	$5.00/$15.00	Qwen-VL	$0.40/$0.40	97%
Fast	GPT-4o-mini	$0.15/$0.60	Qwen-7B	$0.09/$0.09	60%

LangGraph Multi-Model Factory

from enum import Enum
from langchain_openai import ChatOpenAI
import os

class ChineseModel(str, Enum):
    DEEPSEEK_CHAT = "deepseek/deepseek-chat"
    DEEPSEEK_CODER = "deepseek/deepseek-coder"
    QWEN_VL = "qwen/qwen-2-vl-72b-instruct"
    QWEN_FAST = "qwen/qwen-2.5-7b-instruct"
    QWQ_REASONING = "qwen/qwq-32b"
    MOONSHOT_LONG = "moonshot/moonshot-v1-128k"
    AUTO = "openrouter/auto"

def create_llm(model: ChineseModel, **kwargs) -> ChatOpenAI:
    """Factory for OpenRouter LLMs with sensible defaults."""
    return ChatOpenAI(
        model=model.value,
        openai_api_key=os.getenv("OPENROUTER_API_KEY"),
        openai_api_base="https://openrouter.ai/api/v1",
        default_headers={
            "HTTP-Referer": os.getenv("APP_URL", "http://localhost"),
            "X-Title": os.getenv("APP_NAME", "LangChain App")
        },
        **kwargs
    )

# Usage
chat_llm = create_llm(ChineseModel.DEEPSEEK_CHAT)
vision_llm = create_llm(ChineseModel.QWEN_VL)
fast_llm = create_llm(ChineseModel.QWEN_FAST, temperature=0)

Environment Setup

# .env
OPENROUTER_API_KEY=sk-or-v1-...
APP_URL=https://your-app.com      # For attribution (optional)
APP_NAME=Your App Name            # For attribution (optional)

</core_concepts>

For detailed information, see:

• reference/models-catalog.md - Complete model listing with capabilities
• reference/routing-strategies.md - Auto, provider, and custom routing
• reference/langchain-integration.md - LangChain/LangGraph patterns
• reference/cost-optimization.md - Budget management and caching
• reference/tool-calling.md - Function calling patterns
• reference/multimodal.md - Vision, PDF, audio support
• reference/observability.md - Monitoring and tracing

When implementing OpenRouter integration:

• Set OPENROUTER_API_KEY environment variable
• Choose appropriate model for task type (see decision tree)
• Use ChatOpenAI with openai_api_base="https://openrouter.ai/api/v1"
• Add HTTP-Referer and X-Title headers for attribution
• Consider cost implications (Chinese models are 10-100x cheaper)
• Enable streaming for chat interfaces
• Implement fallback chain for production reliability
• Set up cost tracking/budget limits

Emit Outcome Sidecar

As the final step, write to ~/.claude/skill-analytics/last-outcome-openrouter-skill.json:

{"ts":"[UTC ISO8601]","skill":"openrouter-skill","version":"1.0.0","variant":"default",
 "status":"[success|partial|error]","runtime_ms":[estimated ms from start],
 "metrics":{"requests_routed":[n],"models_used":[n],"total_cost_usd":[n]},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Use status "partial" if some stages failed but results were produced. Use "error" only if no output was generated.