Smart Routing

Smart routing automatically selects the best provider for each request based on intent classification. Use claudex run auto to activate it.

How It Works

1. User runs claudex run auto
2. Claude Code sends a request to the proxy
3. The proxy extracts the user’s message and sends it to a local classifier (e.g., Ollama)
4. The classifier returns an intent category (code, analysis, creative, search, math)
5. The routing rules map the intent to the optimal provider profile
6. The request is forwarded to the selected provider

Configuration

[router]
enabled = true
profile = "local-qwen"      # reuse a profile's base_url + api_key for classification
model = "qwen2.5:3b"        # override model (optional, defaults to profile's default_model)

[router.rules]
code = "deepseek"       # coding tasks → DeepSeek
analysis = "grok"       # analytical tasks → Grok
creative = "chatgpt"    # creative writing → ChatGPT
search = "kimi"         # search/research → Kimi
math = "deepseek"       # math/logic → DeepSeek
default = "grok"        # fallback for unclassified

Intent Categories

Intent	Description	Example Tasks
code	Programming and code generation	Write a function, fix a bug, refactor code
analysis	Data analysis and reasoning	Analyze logs, compare approaches, explain concepts
creative	Creative writing and ideation	Write documentation, brainstorm names, draft emails
search	Information retrieval	Find API docs, look up error codes
math	Mathematical and logical tasks	Calculate complexity, prove correctness

Note

Smart routing requires an LLM service for intent classification. It reuses an existing profile’s base_url and api_key. The classifier model should be small and fast — qwen2.5:3b works well.

Usage

# Run with smart routing
claudex run auto

# Override model even with auto-routing
claudex run auto -m grok-3-mini-beta