Most LLM agent frameworks feel like they were designed by a committee - either trying to solve every possible use case with convoluted abstractions or making sure they look great in demos so they can raise millions.

I wanted something lightweight, code-first, and TypeScript-native, so I built a new one. It’s a minimal, strongly-typed agent framework that lets you structure LLM-powered apps without unnecessary complexity.

Key ideas:

Minimal – No unnecessary complexity, just the essentials Code-first – Feels like normal TypeScript, not a magic black box Type-first – Structured inputs/outputs using Zod/@annotations Explicit control – You define how agents behave, no hidden surprises Model-agnostic – OpenAI, Anthropic, DeepSeek, etc.

Links: GitHub: https://github.com/axar-ai/axar Docs: https://axar-ai.gitbook.io/axar

Would love to hear thoughts, especially if you think this approach is flawed.