kalibr

What This Skill Does

Kalibr is an intelligent routing engine for AI agents that enables self-healing production workflows. Unlike simple model proxies that only care about price or uptime, Kalibr learns which specific execution paths—combining models, tools, and parameters—actually perform best for your unique goals. By monitoring telemetry in real-time, Kalibr automatically shifts traffic away from failing or degraded models before your users ever see an error, ensuring that your agents remain resilient without manual intervention or hardcoded fallback logic.

Installation

To integrate Kalibr into your OpenClaw environment, execute the following command in your terminal:

openclaw plugins install @kalibr/openclaw

Ensure you have configured your environment credentials before running your agents:

export KALIBR_API_KEY="your-api-key"
export KALIBR_TENANT_ID="your-tenant-id"

Use Cases

• Reliability Engineering: Automatically route around provider outages (e.g., OpenAI or Anthropic API downtime) without writing nested try/catch blocks.
• Cost Optimization: Experiment with cheaper or faster models while maintaining a baseline success rate, letting Kalibr identify the most efficient path for your specific prompts.
• Agent Performance Tuning: If your agents are performing complex tasks like RAG or multi-step tool use, use Kalibr to compare different tool-use strategies across models to see which consistently delivers accurate results.
• Canary Deployments: Safely test new model configurations on a small fraction of traffic to validate their performance before rolling them out to your entire user base.

Example Prompts

1. "I'm worried about GPT-4o going down during peak hours. How can I use the Kalibr skill to automatically fail over to Claude or Gemini?"
2. "My agent's response quality for data extraction is inconsistent. Can I use Kalibr to test if different temperature settings or models improve the success rate?"
3. "Setup a router for my email extraction task that uses Kalibr to automatically learn which model path yields the best results based on the output validation."

Tips & Limitations

• Warm-up Period: Kalibr requires approximately 20 outcomes to begin making informed routing decisions and 50+ to effectively "lock in" the best paths. Do not expect immediate optimization the moment you deploy.
• Telemetry Matters: The quality of your automated routing depends on accurate success reporting. Ensure your success_when logic or manual router.report calls correctly identify what constitutes a successful agent interaction.
• Continuous Learning: Keep the canary traffic enabled. Kalibr uses 10% of traffic to constantly monitor alternatives, which is crucial for detecting subtle degradations in model performance over time that might not be visible during initial testing.