Kun Workspace
A shared desktop and terminal workspace for taking local project tasks through reviewable, testable delivery.
- Source repo
- KunAgent/Kun
- Stars
- ★ 6.4k
- Last updated
- 1d ago
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · Codex · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A developer fixing a cross-file issue in a local TypeScript project who needs code search, file edits, test execution, and a diff review in one task.
- Not a fit if
- Teams planning commercial distribution, SaaS hosting, or commercial-product integration: separate written authorization is required
- Users wanting only tiny changes without approvals and planning: Agent Graph is experimental and Direct mode fits better
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
Kun is a local-first AI agent workspace for code, writing, design, research, and automation. Its Electron desktop GUI and terminal TUI connect to one local kun serve runtime and share threads, plans, approvals, and task history. The Code workspace supports local-project search and editing, command execution, diffs, tests, and review, while Write and Design can produce documents, exports, HTML prototypes, and design-system material. Its experimental Agent Graph lets a Lead Agent coordinate constrained subagents through dependency-based tasks and explicit handoff checks. It is aimed at people and teams that want local-workspace execution evidence while retaining a choice of model providers.
A user opens a local project in the GUI or TUI and submits a task; both interfaces use the same kun serve runtime. Kun can search and edit files, run Terminal commands, manage Plans and Todos, inspect Git / Worktree state, show inline Diffs and a Changes panel, and use /plan to break down a goal or /review to produce review findings. It accepts image and PDF inputs for research, while Write can export Markdown, HTML, PDF, DOCX, and editable PPTX; Design can create HTML prototypes and DESIGN_SYSTEM.md. Automation uses Schedule, Loop, Hook, MCP, Skills, Extensions, and a local runtime API. Agent Graph creates dependency graphs, dispatches subagents within the parent task's permissions, retains execution history, and supports pause, resume, retry, editing, and stopping.
- A developer fixing a cross-file issue in a local TypeScript project who needs code search, file edits, test execution, and a diff review in one task.
- A technical lead coordinating a verifiable multi-stage change through experimental Agent Graph nodes for research, implementation, and validation.
- A writer turning an outline, source material, or draft into a polished deliverable that can be exported as PDF, DOCX, or editable PPTX.
- A product designer exploring an interface from requirements or reference images, producing an HTML prototype and
DESIGN_SYSTEM.mdfor implementation. - A researcher extracting and organizing evidence from PDFs, images, and web leads into structured conclusions and continuing task context.
- An individual or team documenting a recurring workflow with Schedule, Loop, Hook, MCP, and Skills while keeping recoverable execution records.
How do you install or deploy this agent?
Download the desktop release for macOS (Apple Silicon or Intel), Windows x64, or Linux x64; the desktop package includes the TUI. To run from source:
git clone https://github.com/KunAgent/Kun.git
cd Kun
npm ci
npm run devYou need Node.js 22.19+, npm, and at least one supported model subscription, API, or custom provider. For slower mainland-China network access, use: npm ci --registry=https://registry.npmmirror.com.
How do you use this agent?
At first launch, choose a UI language, sign in to a model subscription or configure an API key, Token Plan, or custom Provider; then open a local project or create a workspace and submit a scoped, verifiable task. Run kun from a project directory to start the bundled TUI, which connects automatically to the same local runtime as the GUI. For source development, use npm run dev:tui for the TUI and npm run typecheck, npm run lint, and npm run test for validation.
What are this agent's strengths and limitations?
- The desktop GUI and terminal TUI share one
kun serveruntime, including threads, plans, approvals, usage, and background tasks. - It keeps local-project file work, Terminal execution, Git / Worktree state, diffs, tests, and
/reviewfindings within one task record. - It supports subscriptions, Coding Plans, Token Plans, APIs, OpenAI Chat Completions / Responses, Anthropic Messages-compatible services, and self-hosted models instead of a single provider.
- Agent Graph adds dependency scheduling, constrained subagents, evidence follow-up, and Lead Agent acceptance for complex tasks, with resumable history.
- Agent Graph is explicitly experimental; its coordination and acceptance flow is better suited to complex work, while Direct mode is faster for simple changes.
- Source use requires Node.js 22.19+, and practical use requires a supported subscription, API, or custom provider configuration.
- Local-first does not mean data never leaves the machine: prompts, attachments, and task context are sent to a selected cloud Provider when one is used.
- Media generation, higher-permission features, and available models depend on the version, operating system, Provider, model capability, and user authorization.
- The README identifies the license as PolyForm Noncommercial 1.0.0; commercial use, distribution, SaaS, hosting, resale, or commercial-product integration requires separate written authorization.
How does this agent compare with similar options?
Compared with a chat interface that only generates answers, Kun is designed to keep requirements, plans, file changes, tool results, tests, review evidence, and delivery in one continuous workflow; Direct mode remains the lighter option for simple tasks.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Kun Workspace This agent | 49 · Major gaps | Desktop appFree + model costs | ★ 6.4k | 1d ago | TypeScript | ChatGPT · Codex · OpenAI API · Claude API |
| Synara | 52 · Major gaps | Desktop appFree + model costs | ★ 2.2k | today | TypeScript | Codex · Claude Code |
| Code Puppy | 57 · Major gaps | CLIFree + model costs | ★ 878 | today | Python | Codex · OpenAI API · Claude API |
| zerostack | 57 · Major gaps | CLIFree + model costs | ★ 1.7k | 2d ago | Rust | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows: README declares local-first, data stored locally by default, and mentions tool approvals, sensitive-action confirmation, sandboxes; SECURITY.md provides vulnerability reporting channels. However, no specific permission model, approval flow details, or data flow diagrams are provided, and publisher identity is unverified. Deductions: least privilege, user confirmation, sensitive data handling are only asserted without implementation details; dependency security has only overrides and audit scripts, no audit results; external effects, rollback, source attribution are only partially documented or unclear.
Evidence shows: README and docs are consistent, package.json scripts are complete, test files demonstrate handling of edge cases (e.g., refusing uncertain operations). However, dependency availability is unverified, and no concrete examples of failure messages are provided. Deductions: self-consistency is good, but dependency availability relies only on package.json declarations, failure messages inferred from tests.
Evidence shows: README clearly identifies target users (developers, writers, designers, researchers) and multiple scenarios (coding, writing, design, research, automation), and states platform support (macOS, Windows, Linux). Capability boundaries are described in docs (e.g., Agent Graph does not expand permissions). However, trigger precision (e.g., commands, shortcuts) is not detailed in provided files. Deductions: audience and scenarios, capability boundaries, environment fit are well documented, but trigger precision evidence is insufficient.
Evidence shows: README provides installation steps, documentation links, FAQ, known limitations (e.g., experimental Agent Graph), license (PolyForm Noncommercial), and version info (package.json version 0.1.0). However, no explicit changelog, and maintenance responsibility only implied via contribution guide and CLA. Deductions: information architecture, install notes, examples and FAQ are sufficient, but naming stability, known limitations, versioning/changelog, maintenance responsibility are insufficiently evidenced.
Evidence shows: README describes outputs as verifiable results (code changes, test results, documents, etc.) and emphasizes value over regular chat boxes. However, cost-benefit is not quantified, and no real usage cases are provided. Deductions: output usability and marginal value are described, but cost-benefit lacks evidence.
Evidence shows: README claims are consistent with docs and test files (e.g., Agent Graph permissions, local data handling), and test files demonstrate verifiable behavior. However, cross-source corroboration is limited, and fact-inference separation is not explicit in docs. Deductions: claim traceability is good, but cross-source corroboration is insufficient, fact-inference separation only partially evident.
- Publisher identity is unverified; verify the source before downloading and installing.
- License is PolyForm Noncommercial; commercial use requires separate authorization. Confirm it fits your use case.
- Dependency security is only asserted via overrides and audit scripts; no actual audit results are provided. Consider running npm audit yourself.
- Agent Graph is marked experimental; use with caution in production.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.4.4 Kun 0.4.4
Release notes excerpt
Kun 0.4.4 0.4.4 让资料、对话与产物更紧密地连在一起：Work 用标签页和分屏串起会话与文件，论文库整合为独立空间；Design 支持分区多画板，数据报告可以直接呈现在对话里并保存到产物库。同时带来工作台新版引导、统一应用连接入口、个人助理任务工作流与存储管理，并改善计划交接、并行任务恢复和 Windows 启动稳定性。 ✨ 新功能 - Work 标签页与论文空间：会话与文件共用标签栏，支持并排对照和会话内预览，论文库统一为独立的检索、阅读与笔记空间。文档选区可直接交给 Kun 改写、扩写或引用，表格选区可查看统计并发起分析。 - 分区多画板设计：同一个 Design 画布可以按分区组织多个页面、状态与方案，继续对话即可追加或修改画板。Claude Code、Codex 及受支持的 ACP Agent 也能在 Code 对话中打开设计，确认方案后再进入代码实现。 - 数据报告与产物预览：对话中可直接展示包含指标、图表和说明的数据报告，并保存到产物库中重新打开和更新。产物卡片增加真实缩略图，查找已有成果更直观。 - 工作台与新版引导：更新圆角工作区、模式切换和可收起的右侧工具条，macOS 侧栏支持通透的玻璃效果。首次使用或从旧版本跨越 0.4.4 升级时提供可跳过的工作台介绍，并可从帮助入口重看各模式引导。 …View on GitHub ↗ -
v0.4.3 Kun 0.4.3
Release notes excerpt
Kun 0.4.3 0.4.3 让计划执行与多 Agent 协作更顺手：在计划卡片中选择执行 Agent，为任务分派填写 Agent 与模型偏好，并在侧栏查看工作 Agent。账户与设置入口进一步整合，新增会员兑换与状态刷新，支持 MiniMax-M3.1-Flash，同时减少计划工作区的磁盘占用，改善工具调用、审批和历史会话的稳定性。 ✨ 新功能 - 计划执行 Agent 选择：在计划卡片中选择负责执行的 Agent，让计划交给合适的引擎完成。 - 任务分派与协作：支持填写 Agent 与模型分派偏好，侧栏在父会话下展示工作 Agent，方便查看协作进展。改进中断子任务的续接；调整分派选择时暂停倒计时，切换的 Agent 会正确应用。 - 账户与会员入口：账户卡片与设置整合到侧栏底部的统一入口，支持会员兑换与会员状态刷新。回到应用时会重新获取会员权益信息。 - 更轻量的计划工作区：计划构建工作树通过写时复制复用已有依赖，减少重复安装与磁盘占用。增加已合并空闲工作树的自动清理，并在可用磁盘空间不足 5 GB 时阻止新建计划构建工作树。 - MiniMax-M3.1-Flash：MiniMax 官方 API 与 Token Plan 模型目录新增 MiniMax-M3.1-Flash。同步适配推理强度选项与用量计价。 …View on GitHub ↗ -
v0.4.2 Kun 0.4.2
Release notes excerpt
Kun 0.4.2 0.4.2 是 0.4.1 之后的增量更新。产物库、可编辑幻灯片和 Office 预览让成果更容易查看与继续修改；交互组件（Gen UI）与 MCP Apps 转为内置能力，Code 工作台、Design 和 Work 写作体验也有了新布局。Changes 面板新增提交与推送，回合撤销只还原 Kun 改过的文件；Agent 名册、协作面板、内置终端、文件树和 Kun TUI 同步升级，启动与空闲资源占用进一步降低。 > 上线状态：个人助手仅在“实验功能”中提供，默认关闭；App / 移动端相关能力仍在预览和内部验证阶段，尚未正式上线。本版本不将它们作为正式功能发布。 ✨ 新功能与体验改进 - 产物库：侧栏新增“产物”入口，集中展示 Kun 在各个项目里做出的设计稿、文档、幻灯片、表格、白板和生成图片。支持网格与列表视图、按类型和项目筛选、搜索与排序；PDF 和白板可直接预览，详情页可查看版本，或回到原会话让 Kun 继续修改。回合结束和窗口重新聚焦时自动刷新。PPTX、DOCX、XLSX 产物打开后可直接在右侧与 Kun 对话修改：当前页、选中文字或单元格区域会作为上下文随消息发送；Kun 改动文件前自动保存版本快照，支持对比与撤销。 …View on GitHub ↗
FAQ
Is Kun locked to one model provider?
Do the GUI and TUI create separate task histories?
kun serve runtime and can share threads, plans, approvals, usage, and background tasks.