0xku/kon

Kon is a minimal coding agent (and also a highly opinionated one)

What it solves

Kon is a minimal coding agent harness designed to provide a lightweight, transparent alternative to bloated agent frameworks. It allows developers to use LLMs for coding tasks without inheriting a massive system prompt or a complex, rigid framework, keeping the context window efficient and the behavior predictable.

How it works

Kon operates as a terminal-based agent with a tiny core system prompt (under 270 tokens) and a small set of built-in tools. It layers project-specific context through external files like AGENTS.md or CLAUDE.md and reusable "skills" (instruction packs) stored in .agents/skills/. It supports a wide range of LLM providers (including GitHub Copilot, OpenAI, DeepSeek, and xAI) and local models via OpenAI-compatible APIs.

Who it’s for

It is built for developers who want a coding agent they can easily read, understand, and adapt, and who prefer a terminal-centric workflow with direct shell access and session management.

Highlights

  • Minimalist Core: Fixed harness stays around 1,000 tokens, reducing prompt bloat.
  • Interactive TUI: Features a terminal UI with fuzzy file searching (@), path completion, and slash commands.
  • Direct Shell Integration: Run shell commands directly (!) or pipe their output to the LLM for analysis (!!).
  • Flexible Context: Loads project guidance from AGENTS.md and supports modular, reusable skills.
  • Session Management: Supports resuming sessions from JSONL files, session compaction to save tokens, and focused handoffs to new sessions.
  • Broad Provider Support: Integrated OAuth and API key support for multiple major AI providers and local model compatibility.

Related

  • Project
    Sentdex/minion

    A lightweight coding agent designed to eliminate context bloat, enabling efficient operation on local LLMs and remote APIs while providing tools to read, write, and execute shell commands.

  • Project
    LiuMengxuan04/MiniCode

    A lightweight terminal coding assistant that implements a Claude Code-like agentic workflow for local development, designed for speed and extensibility.

  • Project
    SWE-agent/mini-swe-agent

    A minimal AI software engineering agent that uses only bash commands to solve coding tasks, providing a high-performance, hackable alternative to complex agent frameworks.

  • Project
    openinterpreter/openinterpreter

    A coding agent that runs in the terminal, optimized for low-cost models through harness emulation and supporting local code execution and computer use.

  • Project
    huggingface/tau

    A readable terminal-based coding agent that can edit code and run commands, designed both as a functional tool and a teaching example for building agentic systems.