Direct answer: A native desktop client for DeepSeek Harness has surfaced via the nightly channel — verified by community download, install and Apple's notarization record as signed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., though DeepSeek has made no official announcement. Official v0.1.7-series release notes now carry extensive desktop entries, confirming the client is wired into the mainline.
What the client looks like
Per Machine Heart's hands-on (September 25): open the app, sign in with a DeepSeek account or an API Key, pick a local folder as the workspace, and assign tasks to the Agent like any desktop AI app. A first-run onboarding flow asks about your use case (office & research, coding, etc.) and preferred workflow — matching the "first-use onboarding" entry in the official v0.1.7-rc.2 changelog.
Four work modes: Standard for most tasks; PTC, where the model writes a small program to batch-organize and invoke tools; Minimal, terminal-only, closest to a classic coding agent; and Create, where the Agent composes plugins, features and modes — consistent with the "Everything is a Plugin" architecture.
Six official plugins are built in: Agent Teams (multi-agent collaboration, experimental), Voice Input (local SenseVoice, experimental), Terminal, Agent Loop, Subagent, and Web Search.
Official signals
- The installer is signed and Apple-notarized under Hangzhou DeepSeek Artificial Intelligence Co., Ltd. — the company entity itself, not a third party.
- Official v0.1.7-rc.2 release notes (verified on GitHub) include desktop entries throughout: first-use onboarding, background running after window close, desktop shortcut management, Windows installer fixes, macOS window-drag fixes and Markdown image preview in desktop chats.
- Still nightly-channel only: no standalone release tag, no announcement. The repo README still recommends
npx @deepseek-ai/dsh webas the standard path. Linux is not supported yet. - The repository has grown to 235k stars as of September 25 (official page), up from 181k on August 22.
How to read it
Until now DSH's recommended entry was a terminal command — distinctly a developer tool. A notarized desktop client with onboarding, work modes and voice input aims squarely at mainstream users. It also slots into the site's Agent Harness glossary and comparison hub: the harness race is no longer about architecture alone but about who lowers the usage floor fastest. Treat it as a candidate build in official distribution, not a stable release — expect rough edges and fast iteration (Machine Heart's fresh install already prompted an update).
Sources
- Machine Heart report (2026-09-25): DeepSeek Harness，终于有官方桌面版了？
- DeepSeek Harness official releases (desktop entries in v0.1.7 series, verified 2026-09-25): github.com/deepseek-ai/deepseek-harness/releases
Direct answer: DeepSeek Harness 的官方桌面客户端已经出现，走 nightly 通道分发——社区下载安装实测可用，安装包经 Apple 公证，签名主体为杭州深度求索人工智能基础技术研究有限公司（DeepSeek 公司主体）；但 DeepSeek 尚未官宣，也不能等同于稳定版发布。官方 v0.1.7 系列更新日志已大量出现桌面端条目，桌面客户端已进入产品主干。
客户端长什么样
据机器之心 9 月 25 日实测：打开 App，登录 DeepSeek 账号或添加 API Key，选一个本地文件夹作为工作区，即可像普通桌面 AI 应用一样直接给 Agent 布置任务。首次使用有引导流程——询问使用方向（办公科研、代码开发等）与偏好工作流，与官方 v0.1.7-rc.2 更新日志里的「桌面端新增首次使用引导」条目相互印证。
四种工作方式：
- 标准模式：适合大多数任务，普通用户最容易上手的 Harness 用法；
- PTC 模式：模型先写一小段程序，再通过程序批量组织调用工具，适合批量处理；
- 极简模式：Agent 只用终端工具完成任务，接近传统 Coding Agent；
- 创造模式：Agent 可编写插件、添加功能或界面，组合工具与提示词创建自己的模式——与「Everything is a Plugin」架构一脉相承。
内置 6 个官方插件：智能体团队（多 Agent 协作，实验性）、语音输入（本机调用 SenseVoice，实验性）、终端、Agent 循环、Subagent、网页搜索。
官方侧信号
- 安装包由 DeepSeek 公司主体对应的 Apple 开发者证书签名，并通过 Apple notarization 公证——不是第三方打包；
- 官方 v0.1.7-rc.2 更新日志（GitHub 实读核对）已大量出现桌面端条目：首次使用引导、关闭窗口后任务后台运行、桌面端快捷键查看与自定义、Windows 安装包启动修复、macOS 窗口拖动修复、桌面端聊天 Markdown 图片预览等；
- 仍走 nightly 通道：无独立 release 标签、无官方公告，仓库 README 推荐启动方式仍是
npx @deepseek-ai/dsh web；暂不支持 Linux； - 仓库 star 数已达 235k（2026-09-25 官方页面实读），较 8 月 22 日的 181k 继续攀升。
怎么看
DSH 此前的推荐入口是一条终端命令，是典型的开发者工具；有了公证签名、新手引导、工作方式选择和语音输入的桌面客户端，瞄准的是主流普通用户。对照站内 Agent Harness 术语页与对比总览：Harness 竞赛从架构之争延伸到「谁先把使用门槛打下来」。定位上应视为已进入官方分发体系的候选构建而非稳定版——功能可能不成熟、迭代很快（机器之心刚装完就收到新版本提示）。
来源
- 机器之心报道（2026-09-25）：DeepSeek Harness，终于有官方桌面版了？
- DeepSeek Harness 官方 releases（v0.1.7 系列桌面端条目，2026-09-25 实读核对）：github.com/deepseek-ai/deepseek-harness/releases