clawd-throttle

What This Skill Does

Clawd Throttle acts as an intelligent traffic controller for your LLM interactions. Instead of relying on a single expensive model for every request, it dynamically evaluates your prompt's complexity across eight dimensions—including token count, code presence, reasoning requirements, and conversation depth—in under 1 millisecond. Based on this analysis, it maps your prompt to the most cost-effective model from a pool of 25+ options across 8 major providers including Anthropic, OpenAI, Google, DeepSeek, xAI, Moonshot, Mistral, and Ollama. By balancing performance against cost, it allows you to stop overpaying for simple tasks like summarization or basic Q&A, while still providing the "heavy lift" models like Opus 4.6 when your prompt legitimately requires it.

Installation

To integrate this skill into your environment, use the OpenClaw package manager:

1. Execute the following in your terminal: clawhub install openclaw/skills/skills/liekzejaws/clawd-throttle
2. Navigate to your installation directory and run npm run setup to initiate the configuration wizard.
3. Provide your preferred API keys for the providers you wish to utilize. While Anthropic or Google are required as a base, you can add others like DeepSeek or xAI to increase your routing flexibility.
4. Select your desired routing mode (eco, standard, or gigachad) to define your base behavior.

Use Cases

Clawd Throttle is ideal for power users and developers who rely on LLMs daily but want to optimize their API spend. It is perfect for teams managing high-volume, automated workflows where simple system prompts are common. It also serves as a "set and forget" solution for developers who don't want to manually toggle models while context-switching between writing code and brainstorming ideas.

Example Prompts

1. "What is the current status of my server deployment?" (The router will identify this as a simple query and route to a low-cost model like Grok 4.1 Fast or Haiku.)
2. "Refactor this function to handle asynchronous streams and add error handling for edge cases." (The router detects code complexity and selects a mid-tier or high-tier model accordingly.)
3. "Analyze the following 5,000-word transcript and provide a 3-bullet point summary." (The router identifies this as a multi-step reasoning task and selects a capable model like Sonnet.)

Tips & Limitations

• Use the set_mode command to quickly pivot if your current tasks become more complex than usual (e.g., switching from "eco" to "gigachad" for a coding sprint).
• Take advantage of manual overrides: type /sonnet or /haiku directly in your prompt to bypass the router for specific edge cases.
• Remember that privacy is a priority: only a SHA-256 hash of your prompt is stored, ensuring your actual data is never sent to the log files. Note that local API performance may vary depending on your network latency to specific provider endpoints.