Kun is a local-first workspace in which an AI agent plans, executes, verifies, and delivers work inside a real project. It organises that work into two main modes: Code, for software delivery, which also includes a Design canvas within the same task, and Work, for writing, document analysis, spreadsheet analysis, and presentation output. The agent reads workspace context, drafts a plan, calls tools, edits files, runs verification, and keeps the resulting evidence attached to the task.
Two interfaces share one runtime. The desktop GUI (macOS, Windows, and Linux) is for observing, reviewing, and controlling a run; the terminal TUI is for keyboard-focused work. Both connect to the same local kun serve process, so threads, goals, plans, approvals, and background tasks are continuous across them rather than split into separate sessions.
A task moves from a stated goal and constraints to a plan, then step-by-step execution, then a review of diffs, tests, reviews, and artifacts before delivery. Requirements and plans can be saved in the project itself, which places them under version control and code review. Sessions, preferences, logs, and runtime data stay on the local machine by default; when you select a cloud model, prompts, attachments, and task context go to that provider, and tool, sensitive-operation, and extension permissions are surfaced in the interface for you to grant or refuse.
Features
- Code mode: project context, file editing, an integrated terminal, Git and worktree handling, diff review, tests, and code review in one workbench
- Design canvas: switch into a design surface within the same Code task to build prototypes and a design system, then carry Design-to-Code context back into implementation
- Work mode: edit Markdown, preview and query PDF and Office documents (Office files stay read-only), analyse spreadsheets, generate presentations from an outline, and sketch ideas on a whiteboard
- Shared local runtime: the GUI and TUI attach to the same kun serve process and see the same threads, plans, approvals, and background tasks
- Evidence-linked tasks: plans, to-dos, tool calls, file changes, browser and terminal results, and approvals are all associated with the task that produced them
- Automation: scheduled tasks, loops, hooks, MCP servers, skills, and installable extensions
- Model providers: presets for ChatGPT and Codex, Claude, Gemini, Cursor, Ollama, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Xiaomi MiMo, plus OpenAI- and Anthropic-compatible services and self-hosted models
- Local-first data: sessions, preferences, and logs are stored on the machine by default, with explicit permission prompts for tool and extension access
