Open sourceTypeScript · ≈285k stars · MIT
DeepSeek Harness
Everything is a Plugin — DeepSeek's open agent harness.
DeepSeek Harness (dsh) is DeepSeek's open-source agent harness, built on an everything-is-a-plugin architecture powered by the Cordis framework. It runs as a local Web UI via npx @deepseek-ai/dsh web, with desktop builds for macOS and Windows. It is a developer preview: every release so far (0.2.x) is a pre-release and the README warns of breaking changes.
Pros & cons
Each point cites its sources and says what kind of evidence backs it. A second reviewer tried to refute every claim before it was kept.
Pros
- DocsIts MIT-licensed "everything is a plugin" architecture (built on Cordis) makes tools, subagents and even the loop replaceable.[1]
- User reportsUsers call it lightweight and fast, and praise two-way subagent messaging that lets parent and child redirect each other mid-task.[1]
- User reportsUsers praise the web UI and the harness's ability to configure itself.[1][2]
- User reportsA user reports you can opt out of your data being used for training.[1]
Cons
- DocsIt is an explicit developer preview whose README warns of compatibility-breaking changes, and users confirm it is "constantly in flux."[1][2]
- User reportsA user found the desktop build enables telemetry by default and attaches session transcripts to DeepSeek API requests, unlike the "dsh web" build.[1]
- User reportsBecause core features are plugins, changing built-in behaviour means keeping downstream patches; one user carries about 25 downstream commits and has written no new plugins.[1]
- MetadataGitHub Issues are disabled on the repo, which makes known bugs hard to track.[1]
- Best for
- Developers who want a fast, hackable web or desktop harness tuned for DeepSeek models and are fine with churn.
- Watch out for
- Turn off desktop telemetry before first launch, and expect breaking changes between releases.
Assessed Oct 29, 2026 · adversarially reviewedSuggest a correction
People
Engineer, DeepSeek
Harness engineering
Activity
- Last push
- · 3d ago
- Latest release
- dsh-v0.2.2-rc.5 (pre-release)
· 3d ago - Last 90 days
- 20+ releases
dsh-v0.2.2-rc.5v0.2.2-rc.5Pre-release
· 3d ago中文 English 体验优化 DeepSeek 模型使用 DeepSeek 服务端搜索,搜索不按次收费;其他模型默认使用 DeepSeek Search API。 @LegGasai Improvements DeepSeek models use DeepSeek server-side search with no per-search charge; other models default to DeepSeek Search API. by @LegGasai Full Changelog:…
dsh-v0.2.2-rc.4v0.2.2-rc.4Pre-release
· 4d ago中文 English 问题修复 优化模型请求收到异常消息响应时的表现。 @turtle2099 修复 LiveShare 在桌面端无法创建分享的问题。 @mektpoy Bug Fixes Improve handling of unexpected model responses. by @turtle2099 Fix LiveShare failing to create shares in the desktop app. by @mektpoy Full Changelog:…
dsh-v0.2.2-rc.3v0.2.2-rc.3Pre-release
· 8d ago中文 English 新增功能 新增实验性「DeepSeek 会话导入」插件,支持将 DeepSeek App 或网页版的历史会话导入工作区,并继续对话。 @imccyu 问题修复 修复 Windows 桌面版应用内更新可能失败、无法自动重新打开的问题。 @winewill New Features Add an experimental DeepSeek chat import plugin to import past conversations from DeepSeek App or its web version into a…
dsh-v0.2.2-rc.2v0.2.2-rc.2Pre-release
· 10d ago中文 English 新增功能 Git Worktree 支持按分支、提交和文件类型审阅改动,并直接处理 GitHub 评审、同步已审阅进度和更新分支。 @LegGasai SubAgent 支持具名与多层协作,AgentTeam 可按会话开启,共用成员与任务面板。 @Dudu-0223, @Yifffan 桌面端支持在设置中切换正式版与开发版更新渠道。 @winewill 体验优化 改善代码审阅时的评论、文件导航与窄屏布局,减少查看大型改动时的卡顿,并放宽大型仓库创建工作树的等待时间。 @LegGasai…
dsh-v0.2.2-rc.1v0.2.2-rc.1Pre-release
· 2w ago中文 English 作为 0.2.2 系列的首个候选版本,本版本汇总了自 v0.2.1-rc.5 以来的主要用户和开发者相关变更。 新增功能 完善手机和平板适配,优化软键盘交互、导航与窄屏布局,保持菜单和操作入口可用。 @Stevring, @turtle2099, @LegGasai, @EricMoin, @Saafo 支持将聊天文本作为引用插入输入框,并在发送后显示紧凑预览。 @lsdsjy, @imccyu 支持将完整历史会话复制为 Markdown,包含可用的原始日志路径。 @tianyicui…
Checked against GitHub on Nov 21, 2026. Release notes are shortened to plain text.
In the news
Launch · Hacker News · Oct 2, 2026
DeepSeek releases a macOS and Windows Desktop app for DeepSeek Harness, its open-source agent harness where everything — tools, sub-agents, even the loop — is a plugin. The repo passed 240k stars within three months of its August launch, making it the fastest-growing harness in the field's history.
In research (9)
Oct 27, 2026 · One of the five coding harnesses compared (54.7%).
Oct 26, 2026 · DSH runs DeepSeek-V4.1-Flash, which sits on the cost-score Pareto frontier.
- A Unified Arena for LLM Agents across Diverse EurogameVerdict: Useful
Oct 26, 2026 · DSH is the native harness for the DeepSeek V4 configurations.
- EmbodiedLab: Benchmarking Embodied Agents That Act to KnowVerdict: Useful
Oct 26, 2026 · Harness for the DeepSeek-V4.1-Flash agent.
- Reasoning Patterns for Efficient AI Coding: An Empirical StudyVerdict: Incremental
Oct 26, 2026 · DeepSeek's harness runs the R1/V4 reasoning models whose traces are studied; early stopping would sit at this layer.
Oct 23, 2026 · The second wrapped harness, which shows the largest reported savings (87.93%).
Showing the 6 newest of 9 papers.
Neighbours on the workbench
TypeScript · ≈395k stars · MIT
openclaw/openclaw
The viral open-source personal AI assistant that really does things.
Maintained by OpenClaw Foundation
v2026.11.6releasedShell · ≈329k stars · MIT
mattpocock/skills
Small, composable agent skills for real engineering work.
Maintained by Matt Pocock (AI Hero)
MDX · ≈323k stars · MIT
obra/superpowers
An agentic skills framework and software methodology that works.
Maintained by Jesse Vincent (obra)
Connections
Relations from the Harness Hub knowledge graph, as charted in the State of Harness reports.
Maintained by · 1
- DeepSeekOrganisation
Mentioned in · 3
- Reconfigurable Action Theories: A Formal Foundation for Dynamically Composable Agent HarnessesPaper · 2026-11-09
- HarnessMM: Recursive Self-Improvement for Harness Token EfficiencyPaper · 2026-10-23
- DeepSeek Harness Desktop for macOS and WindowsNews · 2026-10-02
Sources
Verified Nov 16, 2026
