model-router

Smart Router

> Intelligent model routing for any LLM setup. Automatically routes prompts to the optimal model based on task type, complexity, and model strengths.

Works with: OpenRouter, Alibaba Cloud, Ollama, LM Studio, vLLM, Anthropic Direct, OpenAI Direct, Google AI, and more.

---

Quick Start

1. Choose a Preset

Smart Router includes pre-configured presets for common setups:

| Preset | Models | Best For | Cost | |--------|--------|----------|------| | minimal | 2 models | Simple setups, local LLMs, budget | $ | | alibaba-coding-plan | 3 models | Alibaba Cloud users (free tier) | Free* | | openrouter-all | 7 models | Full OpenRouter access | $$ | | enterprise | 4 models | Production, highest quality | $$$$ |

\* Alibaba Coding Plan: 1200 requests per 5hr shared pool

2. Install

# Via ClawHub (recommended)
clawhub install smart-router

# Or manual
cp -r smart-router/ ~/.openclaw/workspace/skills/

3. Load a Preset

# Copy your chosen preset to active config
cd ~/.openclaw/workspace/skills/smart-router

# Example: Use OpenRouter preset
cp presets/openrouter-all.json config/models.json
cp presets/openrouter-all.json config/categories.json

# Or use the included script
python3 scripts/load-preset.py openrouter-all

4. Customize (Optional)

Edit config/models.json to adjust model IDs for your provider:

{
  "models": {
    "fast": {
      "id": "anthropic/claude-sonnet-4.6",  // Your model ID
      "enabled": true,
      "default": true
    }
  }
}

5. Test

python3 -m python.classifier "fix this bug"
# → {"type": "coding", "model_key": "fast", "model_id": "..."}

---

What This Does

Smart Router analyzes each prompt and routes to the best model:

| If your prompt is... | Route to | Why | |---------------------|----------|-----| | "hi", "thanks", "ping" | Fast model | Low latency, cheap | | "fix this bug", "implement feature" | Coding model | Best SWE-Bench score | | "explain quantum physics" | Reasoning model | Best GPQA score | | "analyze this screenshot" | Vision model | Best multimodal | | "write an email" | Writing model | Best prose quality |

Result: Better quality responses, lower costs, faster responses for simple tasks.

---

Available Presets

`minimal` (2 Models)

Bare minimum setup. Works with any provider.

{
  "fast": "Your fast coding/chat model",
  "smart": "Your smart reasoning/vision model"
}

Example configs:

OpenRouter: claude-sonnet-4.6 + gemini-3.1-pro
Ollama: llama3.1:8b + llama3.1:70b
Alibaba: MiniMax-M2.5 + qwen3.5-plus

`alibaba-coding-plan` (3 Models)

Optimized for Alibaba Cloud's free Coding Plan.

| Model | Use For | |-------|---------| | MiniMax M2.5 | Coding, chat, default | | Qwen3.5-Plus | Science, reasoning, large context | | Kimi K2.5 | Vision, frontend, agent swarms |

`openrouter-all` (7 Models)

Full OpenRouter model selection.

| Model | Use For | Cost/1M | |-------|---------|---------| | GPT-5.3 Codex | Critical coding | $0.75/$3.00 | | Claude Opus 4.6 | Complex reasoning | $15/$75 | | Claude Sonnet 4.6 | Balanced tasks | $3/$15 | | Gemini 3.1 Pro | Vision, video, large context | $0.50/$2.00 | | MiniMax M2.5 | Fast coding | $0.30/$1.20 | | Kimi K2.5 | Frontend, agent swarms | $0.45/$2.20 | | DeepSeek V3.2 | Budget coding | $0.32/$0.89 |

`enterprise` (4 Models)

Premium production setup.

| Model | Use For | |-------|---------| | GPT-5.3 Codex | Default, coding, agentic | | Claude Opus 4.6 | Architecture, writing, analysis | | Gemini 3.1 Pro | Vision, video, multimodal | | Claude Sonnet 4.6 | Simple tasks (cost savings) |

---

Create Your Own Preset

Step 1: Copy a Template

cp presets/minimal.json presets/my-setup.json

Step 2: Edit Models

{
  "name": "My Custom Setup",
  "description": "Personal config",
  "models": {
    "fast": {
      "id": "your-provider/model-id",
      "name": "Friendly Name",
      "provider": "provider-name",
      "context": 100000,
      "strengths": ["coding", "chat"],
      "benchmarks": { "swe-bench": 75 },
      "cost": { "input": 0.50, "output": 1.50 },
      "aliases": ["fast", "default"],
      "enabled": true,
      "default": true
    }
  }
}

Step 3: Edit Routing

{
  "routing": {
    "fallback": "fast",
    "categories": {
      "coding": {
        "model": "fast",
        "priority": 80,
        "keywords": ["code", "fix", "debug", "bug"]
      }
    }
  }
}

Step 4: Load It

python3 scripts/load-preset.py my-setup

---

Model Benchmarks Reference

Data from Artificial Analysis, Hugging Face, and community testing (Feb 2026).

Coding (SWE-Bench Verified)

| Model | Score | Provider | |-------|-------|----------| | GPT-5.3 Codex | 82.1% | OpenAI | | Claude Opus 4.6 | 80.9% | Anthropic | | MiniMax M2.5 | 80.2% | MiniMax | | Gemini 3.1 Pro | 78.5% | Google | | Claude Sonnet 4.6 | 77.2% | Anthropic | | Kimi K2.5 | 76.8% | Moonshot | | Qwen3.5-Plus | 76.4% | Alibaba |

Science & Reasoning (GPQA-Diamond)

| Model | Score | Provider | |-------|-------|----------| | GPT-5.3 Codex | 91.2% | OpenAI | | Claude Opus 4.6 | 89.2% | Anthropic | | Qwen3.5-Plus | 88.4% | Alibaba | | Gemini 3.1 Pro | 87.9% | Google | | Kimi K2.5 | 87.6% | Moonshot |

Vision (MMMU Pro / OCRBench)

| Model | Score | Provider | |-------|-------|----------| | Gemini 3.1 Pro | 78.2% (MMMU) | Google | | Kimi K2.5 | 92.3% (OCRBench) | Moonshot | | GPT-5.3 Codex | 76.8% (MMMU) | OpenAI |

Speed (Tokens/Second)

| Model | Speed | Latency | |-------|-------|---------| | Llama 3.1 8B | 2600 tok/s | ~0.1s | | MiniMax M2.5 | ~80 tok/s | ~0.4s | | Claude Sonnet 4.6 | ~65 tok/s | ~0.6s | | Qwen3.5-Plus | ~50 tok/s | ~0.8s | | Kimi K2.5 | ~35 tok/s | ~1.2s |

---

Configuration Reference

Category Fields

| Field | Type | Description | |-------|------|-------------| | model | string | Model key to route to | | priority | number | Higher = checked first (100 = highest) | | description | string | Human-readable description | | triggers.keywords | array | Match any keyword (case-insensitive) | | triggers.patterns | array | Regex patterns | | triggers.thresholds | object | minTokens, maxWords, minCodeFences, etc. |

Threshold Options

| Threshold | Type | Description | |-----------|------|-------------| | minTokens | number | Minimum estimated tokens (chars / 4) | | maxTokens | number | Maximum estimated tokens | | minWords | number | Minimum word count | | maxWords | number | Maximum word count | | minCodeFences | number | Minimum `` code blocks | | minFilePaths` | number | Minimum file paths detected |

---

Provider-Specific Setup

OpenRouter

1. Get API key: https://openrouter.ai/keys 2. Use openrouter-all or minimal preset 3. Set OPENROUTER_API_KEY env var

Alibaba Cloud

1. Sign up for Coding Plan (free tier) 2. Use alibaba-coding-plan preset 3. Configure API key in gateway config

Ollama (Local)

1. Install: https://ollama.ai 2. Pull models: ollama pull llama3.1:8b llama3.1:70b 3. Use minimal preset with Ollama model IDs

LM Studio (Local)

1. Download: https://lmstudio.ai 2. Load models and start local server 3. Use minimal preset with localhost:1234 endpoints

vLLM (Self-Hosted)

1. Deploy vLLM with your models 2. Use minimal preset with vLLM endpoints 3. Configure base URLs in models.json

---

Debugging

# Enable debug logging
export OPENCLAW_SMART_ROUTER_DEBUG=1
openclaw gateway restart

# Test a prompt
python3 -m python.classifier "your prompt here"

# List available models
python3 -m python.classifier --list

# List available presets
ls presets/

---

File Structure

smart-router/
├── README.md              # This file
├── SKILL.md               # Agent instructions
├── config/
│   ├── models.json        # Active model config
│   └── categories.json    # Active routing rules
├── presets/
│   ├── minimal.json       # 2-model preset
│   ├── alibaba-coding-plan.json
│   ├── openrouter-all.json
│   └── enterprise.json    # Premium preset
├── python/
│   └── classifier.py      # Core Python implementation
├── scripts/
│   └── load-preset.py     # Preset loader script
└── references/
    └── benchmarks.md      # Model benchmark data

---

Requirements

OpenClaw 2026.2+
Python 3.8+ (for Python skill usage)
Model provider configured (any OpenAI-compatible API)

---

Contributing

Add a Preset

1. Create presets/your-preset.json 2. Include 2-10 models with benchmarks 3. Define routing categories 4. Submit PR with benchmark sources

Update Benchmarks

Edit references/benchmarks.md with:

Model name and version
Benchmark scores (SWE-Bench, GPQA, etc.)
Source URL
Date tested

---

License

MIT

---

Support

Issues: https://github.com/ishsharm0/smart-router/issues
Discussions: https://github.com/ishsharm0/smart-router/discussions
Benchmarks: https://artificialanalysis.ai, https://llm-stats.com

Summary

Install to Claude Code