Summary

Kon is a minimal, opinionated coding agent designed for developers seeking a lightweight and understandable tool. It prioritizes a small core prompt and built-in toolset, allowing users to layer project-specific context and configurations as needed. Kon offers a transparent and adaptable experience for users who want to fork, modify, or deeply understand their coding assistant.

How It Works

Kon's core philosophy is minimalism, keeping the default system prompt under 270 tokens and the fixed harness around 1,000 tokens. It externalizes project-specific instructions into AGENTS.md or CLAUDE.md files, loaded dynamically. The agent uses 6 core tools (read, edit, write, bash, grep, find) for fundamental tasks, with optional web search/fetch capabilities. Extensibility is provided via a "skills" system for reusable instruction packs.

Quick Start & Requirements

Install via uv tool install kon-coding-agent or from source (git clone, uv tool install .). Kon targets macOS and Linux; Windows is untested. It requires fd and ripgrep (auto-downloadable). Launch with CLI flags like --provider, --model, --continue, or --resume.

Highlighted Details

• Core Toolset: read, edit, write, bash, grep, find for essential file operations and execution.
• Optional Tools: web_search, web_fetch for external information retrieval.
• Interactive TUI: Features file fuzzy-searching (@), path completion (Tab), queued prompts, steer messages, and slash commands (/model, /resume, etc.).
• Session Management: Sessions stored as JSONL in ~/.kon/sessions/, resumable via CLI or TUI.
• Context Loading: Dynamically loads project guidance from AGENTS.md or CLAUDE.md.
• Extensibility: Supports custom "skills" in .kon/skills/ directories.
• Model Agnostic: Works with OpenAI-compatible endpoints, GitHub Copilot, OpenAI, Azure AI Foundry, and ZhiPu.

Maintenance & Community

The README lacks specific community channels or a roadmap, noting inspiration from pi coding-agent, Amp, and Claude Code.

Licensing & Compatibility

Released under the MIT license, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Explicitly targets macOS and Linux; Windows compatibility is untested.