gitmyhub

Kun

TypeScript ★ 5.4k updated 23h ago

AI agent workspace with Code Write and Design modes built into your application.

A cross-platform desktop app that runs AI agents through a structured clarify-design-plan-code-verify workflow, covering coding, UI design, and long-form writing in one place.

TypeScriptDeepSeekXiaomi MiMoMiniMaxsetup: moderatecomplexity 3/5

Kun is a desktop application that brings AI agents into your daily work through a structured workflow. Instead of starting with a quick chat prompt and hoping for the best, it walks you through a full process: clarify what you need, create design mockups, plan the work, write the code, and then verify the results against the original requirements. It runs on macOS, Windows, and Linux.

The app has three main modes. Code mode connects to your local project folders and can read files, run commands, and edit code while showing you a review panel before any changes are committed. Design mode generates UI mockups, interactive prototypes, and shared design systems from your requirements or text descriptions. Write mode is a standalone workspace for long-form documents with export options to HTML, PDF, and Word formats. All three modes share the same session history, approval settings, and model configuration.

A product manager could use it to draft a feature requirement, generate a visual prototype, turn that into an implementation plan, and hand off to a developer who reviews and executes the plan in Code mode. A solo founder could write product docs, design interfaces, and manage code changes in one place. It also supports connecting to messaging apps like Feishu or WeChat to trigger tasks remotely, and includes a visual workflow builder for automating multi-step agent processes.

What makes this project notable is its choice of AI models. Rather than defaulting to expensive providers, it centers on three Chinese model services — DeepSeek, Xiaomi MiMo, and MiniMax — chosen for their low cost across text, reasoning, vision, voice, image, music, and video generation. The idea is that a longer, multi-step workflow only works if you can afford to run it repeatedly. You can add other providers, but the default setup is designed around keeping the full pipeline cheap enough for everyday use.

The project is licensed for non-commercial use only, meaning it is available for learning and reference but requires separate written permission for any commercial application.

Where it fits