I created yet another coding agent - Its tiny and fun (atleast for me), hope the community finds it useful
Open-source agent uses just 215-token system prompt, runs on consumer hardware like RTX 3090.
A developer has released 'Kon,' a new open-source coding agent that prioritizes minimalism and accessibility over feature bloat. The project distinguishes itself with an extremely lightweight architecture: the system prompt consumes just 215 tokens, with total tool definitions adding up to around 600 tokens, keeping the core overhead under 1,000 tokens before any conversation context is added. This design makes it highly efficient for local execution, as demonstrated by the creator running it using the Zhipu AI GLM-4.7-Flash-Q4 model on a consumer-grade i7-14700F CPU with 64GB RAM and an RTX 3090 GPU with 24GB VRAM.
The project's codebase is intentionally small and comprehensible, containing only 108 files—a stark contrast to popular alternatives like OpenCode (4,107 files) and Pi-Mono (740 files). This minimalism is the core philosophy: Kon is designed to be understood, forked, and extended within a single weekend. While it may lack the extensive model support, test coverage, and surface area of more mature agents, it offers a 'batteries-included' starting point for developers who want a functional coding assistant without navigating a massive codebase.
Available on GitHub and PyPI, Kon represents a growing trend toward streamlined, locally-runnable AI development tools. It takes inspiration from projects like Pi-Coding-Agent but carves its own niche by focusing on developer experience and hackability. For engineers and hobbyists looking to build upon or customize an AI coding assistant without being overwhelmed by complexity, Kon provides a practical and approachable foundation, lowering the barrier to creating personalized AI-powered development environments.
- Extremely lightweight architecture with a 215-token system prompt and under 1k token total overhead before context
- Runs locally on consumer hardware, demonstrated with a GLM-4.7-Flash model on an RTX 3090 GPU
- Minimal 108-file codebase designed to be fully understood and extended in a weekend, unlike competitors with thousands of files
Why It Matters
Lowers the barrier for developers to customize and run efficient, local AI coding assistants without massive codebase complexity.