learn
Your router shouldn't use generic rules; it should learn your specific workflow. hirn_learn provides the tooling to collect, label, and finetune your local classifier.
Data Collection
As you work, the router silently logs every interaction locally. These logs are stored strictly on your local disk at ~/.hirn/logs/ and are never uploaded anywhere.
{"input": "Refactor data layer", "intent": "coding", "files": 12, "chosen_model": "gemini-3.5-flash"}
{"input": "Debug memory leak", "intent": "debugging", "files": 3, "chosen_model": "fable-5"}The Training Loop
Using a function-calling optimized tiny model like Needle, you can quickly adjust the router's behavior to match your personal preferences.
Fire up the local hirn UI to review your historical logs, override misroutes (e.g. telling the router that specific types of PR reviews should always use Fable 5), and build a high-quality personal dataset.
Once you have a curated dataset, a one-click finetuning process teaches Needle to map your specific inputs to the correct routing function call, continuously optimizing your token spend and latency over time.
Continuous Learning
The more you use hirn, the smarter the router gets at predicting exactly how much compute your specific prompts require.
Privacy First
All finetuning data stays locally on your disk. You never leak your prompts, preferences, or codebase to external endpoints.
One-Click Finetune
Quickly train your classifier natively in the UI with a single click, instantly deploying the updated model weights locally.