Harness Desktop
A desktop home for DeepSeek Harness with model routing, plugins, mobile access, and verified updates.
- Source repo
- baiyuscc13724-max/deepseek-harness-desktop
- Stars
- ★ 12
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 58/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specific
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A Windows 10/11 user who wants an installed or portable DeepSeek Harness desktop app while retaining the official workbench as the only main UI.
- Not a fit if
- macOS or Linux users: formal stable support is Windows 10/11 x64 only
- Teams needing vendor endorsement or commercial rights to the Deep Whale artwork, which is non-commercial
- Source review
- 58/100 · Major gaps
What does this agent do, and when should you use it?
Harness Desktop is a community-maintained Electron shell for the official DeepSeek Harness Web UI, which remains the product's sole primary interface and agent core. It starts the local Harness Runtime and adds main-model and subagent routing, a DSH plugin marketplace, managed Skills, an integrated terminal, workspace previews, schedule visibility, themes, and a desktop pet. Users can import files into the workspace uploads/ directory, prepare reviewable edits through the official read and edit tools, and preview common document and media formats in a right-side workspace. Paired mobile devices can reach the current workbench over a LAN, native WebRTC P2P, or an end-to-end encrypted WSS/443 fallback; Android has a signed preview client, while iPhone and iPad use the Safari workbench. Windows 10/11 x64 is the formally supported stable target, while the macOS DMGs remain public previews pending validation on real Apple hardware. It is aimed at users who want DeepSeek Harness's existing interaction and permission model with stronger desktop integration, mobile continuity, and cryptographically verified delivery.
The application launches the official DeepSeek Harness Web UI and local Runtime inside an Electron desktop shell. Its Electron bridge supplies windowing, installation, updating, and desktop features; the Renderer has no Node.js privileges, and the WebView may access only the local Harness Runtime. Providers, models, and credentials are configured in the official Harness settings, and the shell does not keep a second copy of Provider secrets; the main model and subagents may share a model or use separate selections. The DSH marketplace discovers, installs, and updates plugins, while managed Skills use $ triggers and @ remains the file-reference syntax. MCP settings manage stdio and Streamable HTTP connections and retain credential references rather than raw secrets. User-selected files are imported into uploads/, ordinary workspace files can be downloaded, and the official read and edit tools prepare drafts for review; the right-side workspace provides read-only previews for text, source code, HTML, images, audio, video, and PDF files. Schedule monitoring consumes official append-only events with ETag/304 and since-delta refreshes, falling back to one authoritative full replay after a gap, rewind, or generation fork. Desktop delivery uses an Ed25519-signed manifest, HTTPS, declared file sizes, SHA-256 verification, source fallback, component health checks, and automatic rollback. Mobile Sync pairs trusted devices by QR code and carries workbench state and constrained uploads over LAN, native WebRTC P2P, or an end-to-end encrypted WSS/443 relay.
- A Windows 10/11 user who wants an installed or portable DeepSeek Harness desktop app while retaining the official workbench as the only main UI.
- A multi-model user who needs separate model choices for the primary model and subagents while preserving running, resumable, and read-only historical tasks.
- A Harness user who wants in-app DSH plugin management,
$-triggered managed Skills, and@references to workspace files. - Someone who needs to continue a workbench session from Android or a mobile browser, upload constrained documents, and switch between local and encrypted remote connectivity.
- A technical user who wants MCP stdio and Streamable HTTP configuration, bounded terminal access, and workspace file previews in one desktop application.
- A deployment-conscious user who requires signed manifests, SHA-256 package checks, rollback, and preservation of themes, plugins, and routing settings across Harness updates.
How do you install or deploy this agent?
For routine Windows use, download and run the current v1.0.60 stable installer; it installs for the current user and does not require administrator privileges. A portable Windows executable is also provided. Windows 10/11 x64 is the formal stable target; Apple Silicon and Intel macOS DMGs are public previews pending acceptance on real Apple hardware.
Scoop users can install from the project's bucket:
scoop bucket add harness-desktop https://github.com/baiyuscc13724-max/scoop-harness-desktop
scoop install harness-desktop/harness-desktopSource development requires Node.js 24, npm, and Git:
npm install
npm run devRun the documented checks and distribution build before submitting changes or producing packages:
npm run verify
npm run verify:release
npm run distOn first launch, add a provider, model, and the corresponding credentials in the official Harness settings. The desktop shell does not retain a duplicate copy of Provider keys.
How do you use this agent?
- Launch the installed or portable application, then add the chosen provider, model, and required credentials in the official Harness settings.
- Select a workspace and start a new session. Configure the main model and subagents separately when different routing is needed.
- Use the top palette control for desktop themes or open the Deep Whale desktop pet. Plugins, Skills, models, and general preferences remain in the official settings page.
- Explicitly import files into the workspace
uploads/directory and reference them with@. Use the officialreadandedittools to prepare edits for review. - Configure MCP connections as stdio or Streamable HTTP in settings. Starting a local MCP process requires explicit confirmation.
- For mobile use, open Mobile and Remote Sync in settings and scan the pairing code once. LAN connectivity is preferred; for remote use, configure a personal credential-free
wss://endpoint for WebRTC signaling and encrypted relay fallback. The computer must keep Harness Desktop running.
What are this agent's strengths and limitations?
- It preserves the official DeepSeek Harness workbench, sessions, permissions, and agent core instead of creating a second sidebar or duplicate settings system.
- It combines DSH plugins, managed Skills, main-model/subagent routing, MCP management, bounded terminals, and multi-format workspace previews.
- The update chain includes Ed25519 manifest verification, HTTPS on every hop, size and SHA-256 checks, fallback sources, health checks, and automatic rollback.
- Mobile Sync covers LAN, native WebRTC P2P, and end-to-end encrypted WSS/443 fallback with trusted-device pairing.
- The Renderer lacks Node.js privileges, the WebView is restricted to the local Runtime, and external links pass through constrained IPC to the system browser.
- It is community maintained and not endorsed by DeepSeek; core operation depends on the official DeepSeek Harness and its runtime contracts.
- Formal stable support currently centers on Windows 10/11 x64. macOS lacks final validation on real Apple hardware, and no unsigned iOS IPA is distributed.
- Remote mobile use requires the computer to remain online and depends on network availability; a personal WSS relay also requires the user to supply and configure a
wss://endpoint. - Development requires Node.js 24, npm, and Git, while publishing is gated by verification, cloud builds, signing, and cross-source checks.
- The Deep Whale artwork is licensed separately under CC BY-NC-SA 4.0 and cannot be used commercially, even though the desktop shell code uses the MIT License.
- Once authorized, Computer Use can control the entire Windows virtual desktop and does not apply content-level sensitive-action filtering, making the permission scope broad.
How does this agent compare with similar options?
Compared with using the official DeepSeek Harness Web UI alone, this project keeps that UI and agent core but adds Electron packaging, installation and signed updates, plugin and Skills management, model routing, a desktop pet, terminal access, workspace previews, and mobile pairing. On iPhone and iPad, it deliberately uses the live Safari workbench instead of distributing an unsigned IPA that users could not publicly install without an Apple Developer membership. EasyTier and Tailscale are optional mobile connectivity paths; the documented default is LAN first, native WebRTC P2P for remote connections, and encrypted WSS/443 fallback.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Harness Desktop This agent | 58 · Major gaps | Desktop appFree + model costs | ★ 12 | 2mo ago | JavaScript | — |
| DeepSeek Bot | 72 · Some gaps | Desktop appFree + model costs | ★ 6.3k | today | JavaScript | — |
| DeepChat | 70 · Some gaps | Desktop appFree + model costs | ★ 6.4k | today | TypeScript | OpenAI API · Claude API |
| Hope Agent | 68 · Some gaps | Desktop appFree + model costs | ★ 1.9k | 1d ago | Rust | Codex · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md, the README, and workflow notes describe isolation, signed updates, credential boundaries, authorization, and rollback, supporting ordinary safeguards and earning mostly 2s. Recovery paths for cache cleanup, component updates, and sync are described in detail, so rollback earns 3. Deductions: SECURITY.md permits persistent authorization to control the entire Windows virtual desktop without per-action confirmation or content-level filtering. The README also describes Computer Use and MCP, but the supplied code samples do not verify their enforcement boundaries, so confirmation and external effects score 1. Dependency security scores 1 because the supplied audit reports concern older versions and do not establish the current candidate's dependency state. Attribution scores 2: upstream assets and licensing are identified, but publisher identity and maintenance ownership remain limited.
The README, package.json, workflows, and security reviews describe pinned dependencies, gates, error handling, and release rollback, supporting scores of 2 for dependency availability and failure messages. Self-consistency scores 1: the README identifies 1.0.61 as a candidate with 1.0.60 stable assets, while the supplied security reviews cover 1.0.26–1.0.30 and describe different upstream Runtime versions. These materials do not form a consistent validation record for the current candidate.
The README covers desktop, mobile, development, and collaboration scenarios, and distinguishes stable releases from previews; audience coverage and environment fit score 2. Capability boundaries and trigger conditions are also described, including team activation and degradation when dependencies are unavailable, earning 2. Deductions reflect that many boundaries are assertions in prose, the supplied source samples cover only a small part of the product, and suitability guidance is incomplete.
The README provides downloads, installation steps, a quick start, and links to architecture and security documentation, supporting scores of 2 for information architecture and installation notes. The MIT license is complete and scores 3. Product boundaries and preview limitations are stated, earning 2 for known limitations. Naming stability and changelog/versioning score 1: the package version is clear, but supplied reviews are substantially older and do not establish a verifiable current change history. Examples and FAQ score 1: quick-start commands exist, but no FAQ is provided. Maintenance responsibility scores 2: the project identifies itself as community-maintained and unofficial and provides a security reporting route, but names no specific maintainer or response commitment.
The product has a clear purpose and combines the workbench, plugins, routing, desktop features, and mobile connection while retaining the official workbench as its main interface; output usability and marginal value score 2. Cost-benefit scores 1 because the README describes a broad feature set with substantial complexity, while the supplied material does not establish the practical benefit or cost for ordinary users.
The README, security policy, package.json, release workflows, and security reviews make many implementation and gate claims traceable, earning 2. Cross-source corroboration scores 1 because security descriptions differ across the README, SECURITY.md, and reviews for different versions, while the supplied source samples cover little implementation. Fact/inference separation scores 2: the documents often label previews, pending release checks, and unsigned-package limits; these statements do not replace verification that was not performed in this static review.
- Computer Use can retain authorization to control the entire Windows virtual desktop; review the trust scope before enabling it.
- The supplied older security reviews do not corroborate the current candidate's runtime version or release validation. Do not treat historical validation statements as completed verification for 1.0.61.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/deepseek-harness-desktop)<a href="https://followagents.com/en/agents/deepseek-harness-desktop"><img src="https://followagents.com/badge/deepseek-harness-desktop.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.0.61 Harness Desktop v1.0.61
Release notes excerpt
Harness Desktop 1.0.61 上一稳定版是不可变的 v1.0.60。现行官方 Harness Runtime 与依赖图继续固定 0.1.2-alpha.5，本次不升级官方依赖。本说明描述候选源码；发布完成状态以唯一发布器核验的公开资产及签名更新源为准。 本次更新 - 纳入 PR #47：运行时从安全动态端口范围选择独立回环端口，避免 Chromium 拒绝端口导致工作台打不开；团队 SSE 首次握手有界等待，停滞连接会释放。没有放宽 Chromium 安全策略。 - 手机连接与同步：WSS 旧连接的错误/关闭回调不能破坏新代连接；工作区清单 HTTP/WS 鉴权统一、取消贯穿准备与请求阶段，关闭过程保留错误吸收并有界结束。 - 手机恢复与交互：真实连接提示、恢复失败及手动重试、附件预览焦点和可访问性同步 Android/iOS；不因失败自动重发用户消息。 - 手机控制安全：命令和结果继续绑定确切设备连接及回执身份，拒绝旧连接结果，不新增权限或绕过确认。 范围与状态 …View on GitHub ↗ -
v1.0.60 Harness Desktop v1.0.60
Release notes excerpt
Harness Desktop 1.0.60 本次仅修复代理团队的范围管理、正常等待和停止后清理，不包含已经取消的浏览器、离屏或原生预览任务。官方 Harness 依赖继续固定在 0.1.2-alpha.5，不作新的运行时升级。 上一稳定版是不可变的 v1.0.59。本说明描述候选源码；发布是否完成以唯一发布器核验的公开资产与签名更新源为准。 三项修复 - 范围提醒：显示首次成员开工时的内部任务数与之后新增的任务数，提醒对照原始目标及用户最新要求收口。任务数量不是用户需求数量，也不是授权证明；历史时间不完整时不编造基线。 - 静默等待：team_status.wait_for_change=true 在当前已授权回合内订阅真实进展，而不是不断查询、生成空轮次或汇报“仍在等待”。Stop、取消及插件退出会释放订阅；没有可继续产生成果的成员时返回需处理状态。它不操作 Goal、不补发权限、不启动成员，也不替代自动接力的授权边界。 - 停止后直接清理：精确根会话在直接用户回合内可取消已暂停任务、关闭已停止团队，不再要求先 Resume。取消仍受任务 revision、pause epoch 和审计历史约束；成员只被排空，不发送恢复或退休工作提示。排空失败仍保持暂停，遇到更新的 Stop 拒绝旧清理结果。 版本与验证 …View on GitHub ↗ -
v1.0.59 Harness Desktop v1.0.59
Release notes excerpt
Harness Desktop 1.0.59 v1.0.59 候选将官方 Harness 完整依赖图升级到 0.1.2-alpha.5，并集中解决长时间、多团队、多设备运行中已经复现的可靠性与性能问题：Agent Teams 的全局自动接力、计划重提交、admission/recovery/retirement、实时状态和 Unicode 工作区边界得到收敛；长会话滚动、手机编辑器与导航恢复当前官方合同；Mobile Sync、Schedule、设备预览、缓存维护和团队投影改为可回滚、可验证的增量路径。 上一稳定版是不可变的 v1.0.58。其 Tag、Release 资产、签名 Android APK、组件、release-manifest.json、镜像与 stable feeds 均不移动、不覆盖。本文件描述 v1.0.59 候选源码与发布要求，不代表新版本已经公开、已经签名或已经进入 stable feed。 官方 Harness 0.1.2-alpha.5 …View on GitHub ↗
FAQ
Is this an official DeepSeek application?
Does installation require administrator access or duplicate my model credentials?
Are macOS and mobile clients production-supported?
What happens when direct mobile connectivity fails?
Does the desktop pet inspect conversations or the screen?
Thanks — we use this to improve the page.