llm-router（LLM router 搜索）

LLM Router

Overview

Route AI requests to different LLM providers using SwiftOpenAI-CLI's agent mode. This skill automatically configures the CLI to use the requested provider (OpenAI, Grok, Groq, DeepSeek, or OpenRouter), ensures the tool is installed and up-to-date, and executes one-shot agentic tasks.

Core Workflow

When a user requests to use a specific LLM provider (e.g., "use grok to explain quantum computing"), follow this workflow:

Step 1: Ensure SwiftOpenAI-CLI is Ready

Check if SwiftOpenAI-CLI is installed and up-to-date:

scripts/check_install_cli.sh

This script will:

• Check if swiftopenai is installed
• Verify the version (minimum 1.4.4)
• Install or update if necessary
• Report the current installation status

Step 2: Configure the Provider

Based on the user's request, identify the target provider and configure SwiftOpenAI-CLI:

scripts/configure_provider.sh <provider> [model]

Supported providers:

• openai - OpenAI (GPT-4, GPT-5, etc.)
• grok - xAI Grok models
• groq - Groq (Llama, Mixtral, etc.)
• deepseek - DeepSeek models
• openrouter - OpenRouter (300+ models)

Examples:

# Configure for Grok
scripts/configure_provider.sh grok grok-4-0709

# Configure for Groq with Llama
scripts/configure_provider.sh groq llama-3.3-70b-versatile

# Configure for DeepSeek Reasoner
scripts/configure_provider.sh deepseek deepseek-reasoner

# Configure for OpenAI GPT-5
scripts/configure_provider.sh openai gpt-5

The script automatically:

• Sets the provider configuration
• Sets the appropriate base URL
• Sets the default model
• Provides guidance on API key configuration

Step 3: Verify API Key

The configuration script automatically checks if an API key is set and will stop with clear instructions if no API key is found.

If API key is missing:

The script exits with error code 1 and displays:

• ⚠️ Warning that API key is not set
• Instructions for setting via environment variable
• Instructions for setting via config (persistent)

Do not proceed to Step 4 if the configuration script fails due to missing API key.

Instead, inform the user they need to set their API key first:

Option 1 - Environment variable (session only):

export XAI_API_KEY=xai-...           # for Grok
export GROQ_API_KEY=gsk_...          # for Groq
export DEEPSEEK_API_KEY=sk-...       # for DeepSeek
export OPENROUTER_API_KEY=sk-or-...  # for OpenRouter
export OPENAI_API_KEY=sk-...         # for OpenAI

Option 2 - Config file (persistent):

swiftopenai config set api-key <api-key-value>

After the user sets their API key, re-run the configuration script to verify.

Step 4: Execute the Agentic Task

Run the user's request using agent mode:

swiftopenai agent "<user's question or task>"

Agent mode features:

• One-shot task execution
• Built-in tool calling
• MCP (Model Context Protocol) integration support
• Conversation memory with session IDs
• Multiple output formats

Examples:

# Simple question
swiftopenai agent "What is quantum entanglement?"

# With specific model override
swiftopenai agent "Write a Python function" --model grok-3

# With session for conversation continuity
swiftopenai agent "Remember my name is Alice" --session-id chat-123
swiftopenai agent "What's my name?" --session-id chat-123

# With MCP tools (filesystem example)
swiftopenai agent "Read the README.md file" \
  --mcp-servers filesystem \
  --allowed-tools "mcp__filesystem__*"

Usage Patterns

Pattern 1: Simple Provider Routing

User Request: "Use grok to explain quantum computing"

Execution:

# 1. Check CLI installation
scripts/check_install_cli.sh

# 2. Configure for Grok
scripts/configure_provider.sh grok grok-4-0709

# 3. Execute the task
swiftopenai agent "Explain quantum computing"

Pattern 2: Specific Model Selection

User Request: "Ask DeepSeek Reasoner to solve this math problem step by step"

Execution:

# 1. Check CLI installation
scripts/check_install_cli.sh

# 2. Configure for DeepSeek with Reasoner model
scripts/configure_provider.sh deepseek deepseek-reasoner

# 3. Execute with explicit model
swiftopenai agent "Solve x^2 + 5x + 6 = 0 step by step" --model deepseek-reasoner

Pattern 3: Fast Inference with Groq

User Request: "Use groq to generate code quickly"

Execution:

# 1. Check CLI installation
scripts/check_install_cli.sh

# 2. Configure for Groq (known for fast inference)
scripts/configure_provider.sh groq llama-3.3-70b-versatile

# 3. Execute the task
swiftopenai agent "Write a function to calculate fibonacci numbers"

Pattern 4: Access Multiple Models via OpenRouter

User Request: "Use OpenRouter to access Claude"

Execution:

# 1. Check CLI installation
scripts/check_install_cli.sh

# 2. Configure for OpenRouter
scripts/configure_provider.sh openrouter anthropic/claude-3.5-sonnet

# 3. Execute with Claude via OpenRouter
swiftopenai agent "Explain the benefits of functional programming"

Provider-Specific Considerations

OpenAI (GPT-5 Models)

GPT-5 models support advanced parameters:

# Minimal reasoning for fast coding tasks
swiftopenai agent "Write a sort function" \
  --model gpt-5 \
  --reasoning minimal \
  --verbose low

# High reasoning for complex problems
swiftopenai agent "Explain quantum mechanics" \
  --model gpt-5 \
  --reasoning high \
  --verbose high

Verbosity levels: low, medium, high Reasoning effort: minimal, low, medium, high

Grok (xAI)

Grok models are optimized for real-time information and coding:

• grok-4-0709 - Latest with enhanced reasoning
• grok-3 - General purpose
• grok-code-fast-1 - Optimized for code generation

Groq

Known for ultra-fast inference with open-source models:

• llama-3.3-70b-versatile - Best general purpose
• mixtral-8x7b-32768 - Mixture of experts

DeepSeek

Specialized in reasoning and coding:

• deepseek-reasoner - Advanced step-by-step reasoning
• deepseek-coder - Coding specialist
• deepseek-chat - General chat

OpenRouter

Provides access to 300+ models:

• Anthropic Claude models
• OpenAI models
• Google Gemini models
• Meta Llama models
• And many more

API Key Management

Recommended: Use Environment Variables for Multiple Providers

The best practice for using multiple providers is to set all API keys as environment variables. This allows seamless switching between providers without reconfiguring keys.

Add to your shell profile (~/.zshrc or ~/.bashrc):

# API Keys for LLM Providers
export OPENAI_API_KEY=sk-...
export XAI_API_KEY=xai-...
export GROQ_API_KEY=gsk_...
export DEEPSEEK_API_KEY=sk-...
export OPENROUTER_API_KEY=sk-or-v1-...

After adding these, reload your shell:

source ~/.zshrc  # or source ~/.bashrc

How it works:

• SwiftOpenAI-CLI automatically uses the correct provider-specific key based on the configured provider
• When you switch to Grok, it uses XAI_API_KEY
• When you switch to OpenAI, it uses OPENAI_API_KEY
• No need to reconfigure keys each time

Alternative: Single API Key via Config (Not Recommended for Multiple Providers)

If you only use one provider, you can store the key in the config file:

swiftopenai config set api-key <your-key>

Limitation: The config file only stores ONE api-key. If you switch providers, you'd need to reconfigure the key each time.

Checking Current API Key

# View current configuration (API key is masked)
swiftopenai config list

# Get specific API key setting
swiftopenai config get api-key

Priority: Provider-specific environment variables take precedence over config file settings.

Advanced Features

Interactive Configuration

For complex setups, use the interactive wizard:

swiftopenai config setup

This launches a guided setup that walks through:

• Provider selection
• API key entry
• Model selection
• Debug mode configuration
• Base URL setup (if needed)

Session Management

Maintain conversation context across multiple requests:

# Start a session
swiftopenai agent "My project is a React app" --session-id project-123

# Continue the session
swiftopenai agent "What framework did I mention?" --session-id project-123

MCP Tool Integration

Connect to external services via Model Context Protocol:

# With GitHub MCP
swiftopenai agent "List my repos" \
  --mcp-servers github \
  --allowed-tools "mcp__github__*"

# With filesystem MCP
swiftopenai agent "Read package.json and explain dependencies" \
  --mcp-servers filesystem \
  --allowed-tools "mcp__filesystem__*"

# Multiple MCP servers
swiftopenai agent "Complex task" \
  --mcp-servers github,filesystem,postgres \
  --allowed-tools "mcp__*"

Output Formats

Control how results are presented:

# Plain text (default)
swiftopenai agent "Calculate 5 + 3" --output-format plain

# Structured JSON
swiftopenai agent "List 3 colors" --output-format json

# Streaming JSON events (Claude SDK style)
swiftopenai agent "Analyze data" --output-format stream-json

Troubleshooting

Common Issues

Issue: "swiftopenai: command not found"

Solution: Run the check_install_cli.sh script, which will install the CLI automatically.

Issue: Authentication errors

Solution: Verify the correct API key is set for the provider:

# Check current config
swiftopenai config list

# Set the appropriate API key
swiftopenai config set api-key <your-key>

# Or use environment variable
export XAI_API_KEY=xai-...  # for Grok

Issue: Model not available

Solution: Verify the model name matches the provider's available models. Check references/providers.md for correct model names or run:

swiftopenai models

Issue: Rate limiting or quota errors

Solution: These are provider-specific limits. Consider:

• Using a different model tier
• Switching to a different provider temporarily
• Checking your API usage dashboard

Debug Mode

Enable debug mode to see detailed HTTP information:

swiftopenai config set debug true

This shows:

• HTTP status codes and headers
• API request details
• Response metadata

Resources

This skill includes bundled resources to support LLM routing:

scripts/

• check_install_cli.sh - Ensures SwiftOpenAI-CLI is installed and up-to-date
• configure_provider.sh - Configures the CLI for a specific provider

references/

• providers.md - Comprehensive reference on all supported providers, models, configurations, and capabilities

Best Practices

1. Always check installation first - Run check_install_cli.sh before routing requests
2. Configure provider explicitly - Use configure_provider.sh to ensure correct setup
3. Verify API keys - Check that the appropriate API key is set for the target provider
4. Choose the right model - Match the model to the task (coding, reasoning, general chat)
5. Use sessions for continuity - Leverage --session-id for multi-turn conversations
6. Enable debug mode for troubleshooting - When issues arise, debug mode provides valuable insights
7. Reference provider documentation - Consult references/providers.md for detailed provider information

Examples

Example 1: Routing to Grok for Real-Time Information

# User: "Use grok to tell me about recent AI developments"

scripts/check_install_cli.sh
scripts/configure_provider.sh grok grok-4-0709
swiftopenai agent "Tell me about recent AI developments"

Example 2: Using DeepSeek for Step-by-Step Reasoning

# User: "Ask deepseek to explain how to solve this algorithm problem"

scripts/check_install_cli.sh
scripts/configure_provider.sh deepseek deepseek-reasoner
swiftopenai agent "Explain step by step how to implement quicksort"

Example 3: Fast Code Generation with Groq

# User: "Use groq to quickly generate a REST API"

scripts/check_install_cli.sh
scripts/configure_provider.sh groq llama-3.3-70b-versatile
swiftopenai agent "Generate a REST API with authentication in Python"

Example 4: Accessing Claude via OpenRouter

# User: "Use openrouter to access claude and write documentation"

scripts/check_install_cli.sh
scripts/configure_provider.sh openrouter anthropic/claude-3.5-sonnet
swiftopenai agent "Write comprehensive documentation for a todo app API"

Example 5: GPT-5 with Custom Parameters

# User: "Use gpt-5 with high reasoning to solve this complex problem"

scripts/check_install_cli.sh
scripts/configure_provider.sh openai gpt-5
swiftopenai agent "Design a distributed caching system" \
  --model gpt-5 \
  --reasoning high \
  --verbose high