Mukt started as a practical challenge: keep an AI assistant useful while minimizing model spend. The core idea was to automatically discover currently available free models on OpenRouter and route tasks to the best fit at runtime.
Why this project exists
Most agent demos ignore cost constraints. In real products, cost controls are part of the architecture. Mukt was designed to keep response quality acceptable while driving token cost as close to zero as possible for common workloads.
What I implemented
- Automatic discovery of free and available models
- Capability-aware routing (chat, structured output, tool-call suitability)
- Function-calling support for tool use workflows
- Fallback logic when a selected provider becomes unavailable
Tradeoffs and outcomes
The routing layer adds complexity, but it makes the assistant resilient and cost-aware. The result is a lightweight Vanilla JS agent that stays usable in production-like conditions, not only in happy-path demos.
Stack
Vanilla JS, OpenRouter APIs, function calling/tool-use patterns.