Kun

Kun is a local-first AI agent workbench with a desktop GUI and a terminal TUI. Both interfaces connect to the same local runtime and share threads, goals, plans, approvals, subagents, and background tasks, so a task started in one continues in the other. Agents read the workspace context, make a plan, call tools, edit files, run checks, and keep the evidence (plans, tool calls, file changes, terminal and browser output, approvals) attached to the task.

Work is organized into modes. Code covers software delivery with project context, file editing, a terminal, Git and worktrees, diffs, tests, and review, plus a Design canvas inside the same task for prototypes, design systems, and design-to-code context. Work covers writing and document tasks: editing Markdown, previewing and analyzing PDF and Office files (kept read-only), analyzing spreadsheets, building presentations from an outline, and sketching on a whiteboard. Rooms puts several agents in one chat-style space with private and group conversations, a coordinator or peer collaboration mode, and proposal cards that only the person can accept. Each execution task in a room runs in its own Git worktree and goes to a reviewer agent before delivery.

Sessions, preferences, logs, and runtime data stay on the local machine by default, and prompts and context go to a cloud provider only when one is selected. Kun is not tied to one model: presets cover ChatGPT and Codex, Claude, Gemini, Cursor, Ollama, DeepSeek, Kimi, GLM, Qwen, MiniMax, and Xiaomi MiMo, and subscriptions, API keys, OpenAI or Anthropic compatible services, and self-hosted models connect through provider configuration. It runs on macOS, Windows, and Linux.