Kun makes a bet most coding-agent apps avoid: that AI coding should begin with a written requirement, not a one-line prompt, and that this only holds up if the model bill stays cheap enough to run the whole loop dozens of times a day. The two ideas are connected on purpose, and the second is what makes the first survive as a daily habit.
Kun is a desktop workspace built on Electron and React, running on macOS, Windows and Linux. Instead of handing an agent a sentence and letting it edit code directly, you start by drafting a requirement. A requirement AI helps you fill in the open questions, do pre-implementation research and mark out the boundaries. That draft gets saved as .kunsdd/draft/.../requirement.md, with structured requirement blocks, acceptance criteria and a history you can trace. From there you move into design, then into a plan through /plan and create_plan, which writes a GUI-managed .kunsdd/plan/... file whose steps stay linked to the requirement. When the requirement changes, Kun can prompt you to re-plan so the two do not drift apart.
Three modes over one runtime
The workflow runs across three parallel modes. Code faces a real codebase: it reads project context, runs shell commands, edits files, and shows every change in an inline diff and a review panel before you commit. Design turns requirement fragments into UI mockups, infographics, interactive HTML prototypes and a shared DESIGN_SYSTEM.md, then hands the result to Code for implementation. Write is a standalone Markdown workspace with a file tree, Live, Source, Split and Preview views, selection-based inline editing, and export to HTML, PDF, DOC and DOCX. From a Markdown file you can also trigger PPT Master, which confirms an outline, page count, audience and visual direction with you first, then generates a native, editable .pptx into a presentations/ directory without touching the source file. That step needs Python 3.10 or newer on your machine.
All three modes share one thing: the same Kun runtime, provider configuration, approval mechanism and session handling. Kun ships its own local runtime, reached through kun serve, which exposes an HTTP and SSE boundary and runs a cache-first agent loop with an append-only event log, usage stats and context compression. Sessions, logs, preferences and runtime config stay on your machine by default, and model calls go out through your own provider credentials.
Why the model choice is the design
Kun does not treat its default models as ordinary optional providers. It ships around three Chinese cost-effective vendors: DeepSeek as the main text and reasoning model, Xiaomi MiMo for long-context text, vision input, speech-to-text and text-to-speech, and MiniMax for image, voice, music and video generation. The stated reason is economic. A requirements-first flow produces more rounds of clarification, research, structuring, planning, execution and validation than plain chat, and it leans hard on repeated model calls. If those calls cost too much, the whole approach never becomes something you reach for every day. So Kun organizes its first-run setup and settings around these three, routing light clarification to fast models and heavy code and reasoning to stronger ones. You can still add OpenAI-compatible, self-hosted or custom providers, but the default experience is built around keeping the running cost low.
The rest of the surface follows the same continuous-workflow idea: a visual "Create Loop" canvas for node-based orchestration in the style of n8n or dify, remote entry points through Feishu, Lark and WeChat plus local webhooks and scheduled tasks, MCP servers and Skills, and a .kunx extension platform with an SDK and Extension API v1. One caution worth naming: Kun is released under the PolyForm Noncommercial License 1.0.0, so it is for study and reference, and any commercial use needs separate written permission from the author.
Kun is young, created in May 2026, and the README reads more like a product manifesto than a track record. What it lays out clearly is a single coherent argument: tie requirements to plans to code to review, and pick the models that let you actually afford to run that chain.
GitHub: https://github.com/KunAgent/Kun
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)