multi-llm

What This Skill Does

The Multi-LLM skill introduces a dynamic routing system to the OpenClaw AI agent, allowing users to switch from the default Claude Opus 4.5 model to specialized local models based on the specific task requirements. By prefixing prompts with the 'multi llm' command, the agent analyzes the input's context and keywords to automatically select the most suitable local engine for coding, reasoning, translation, or general inquiries. This enables significant cost efficiency and performance optimization, particularly when handling resource-intensive tasks like logic heavy computations or long-form code generation.

Installation

To install this skill, run the following command within your OpenClaw interface: clawhub install openclaw/skills/skills/leohan123123/mlti-llm-fallback Ensure you have the necessary system resources (RAM/VRAM) to support the specific local models (e.g., 42GB for deepseek-r1:70b) as specified in the model mapping documentation. The system automatically manages the fallback chain if a primary model instance fails to load.

Use Cases

This skill is ideal for power users who need domain-specific expertise. Use the coding mode for complex refactoring projects, the reasoning mode for mathematical proofs or high-level logical analysis, and the Chinese mode for rapid localization tasks. It removes the guesswork of model selection, ensuring that your task is always mapped to the engine that yields the highest accuracy.

Example Prompts

1. "multi llm Please debug this asynchronous Python function and suggest a performance refactor."
2. "multi llm Analyze this mathematical proof for the Riemann hypothesis and explain the logic step-by-step."
3. "multi llm Translate these technical notes into Chinese and summarize the key findings."

Tips & Limitations

• Default Behavior: If you do not include 'multi llm', you are always using the high-performance Claude Opus 4.5 model.
• Fallback Support: Every category features a robust fallback chain. If your hardware cannot support the top-tier model, the agent will automatically downgrade to the next available size in the chain.
• Customization: You can force specific categories by appending them to the trigger, such as 'multi llm coding' to override the detection logic entirely.
• Resource Usage: Always monitor your local memory usage when triggering large models like deepseek-r1:70b to prevent system instability.