DeepSeek Harness Desktop: a Windows GUI for the DSH runtime
Open-source Windows desktop client and GUI for DeepSeek Harness — zero-setup installer with Codex, plugins, skills, SSH, mobile remote access, and 11 skins.
At a glance
- What is it?
- A community-built Electron client that wraps the DeepSeek Harness Web UI into an installer for Windows 10 and 11, with macOS arm64 and Linux x64 previews. It ships the runtime, adds plugins, skills, SSH and remote control, and pins one specific DSH alpha release.
- Who is it for?
- Adopt it if you want the DSH runtime, a Web UI, plugins and skills on a Windows 10 or 11 x64 machine without assembling Node, Git and pnpm yourself, and if you accept that the desktop host tracks a pinned alpha of the runtime. Skip it if you need a vendor-supported client: the README states this is not the official DeepSeek desktop product, and it deliberately avoids the official application identity, data directories and update source.
- Can I use it commercially?
- Yes. BSD-3-Clause is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 2 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap DeepSeek Harness Desktop fills
DeepSeek Harness is a runtime and a web surface. Getting it onto a developer laptop normally means installing Node.js, a package manager, Git and the runtime itself, then keeping those versions aligned. The README states plainly that DeepSeek has not released a standalone desktop product, and that this project fuses the public DSH Runtime and Web UI with the desktop capabilities described in the official repository into something you can install.
The target reader is a Windows developer who wants an agent workbench without a setup session. The README says the installer already contains the main runtime components, so Node.js, Git and a separate DSH install are not prerequisites. That is the whole pitch, and it is a narrow one: you are trading control over your toolchain for a bundled environment that the desktop host manages.
It is explicitly not an official client. The README says the project does not use or overwrite the application identity, data directory or update source of any future official desktop build, and it runs under its own desktop profile so an existing DSH configuration is not overwritten. If you already have a carefully tuned DSH Home, that separation is the detail worth checking first.
How the desktop host relates to the DSH runtime
The application is an Electron shell around the DeepSeek Harness Web Surface. The desktop host starts and supervises a local DSH Runtime, and the window renders the web UI that talks to it. Plugins, their installation, updates and runtime lifecycle are managed by the host rather than by the user.
That layering explains most of the project's behaviour. Because the host owns the runtime, it can pin an exact version: the 4.1.0 notes say the build locks @deepseek-ai/dsh@0.1.6-alpha.1 and migrates Agent, Session, PTC, Workflow, Sandbox and Agent Team to the new contracts. It also means the host must adapt whenever the runtime changes shape, which is why a runtime bump is treated as a release event rather than a dependency refresh.
The host also owns state. Session and long-context data, workspaces, models, credential references, skills, plugins, skins, the desktop pet and collaboration settings are backed up before modification, with rollback on failure, according to the 4.1.0 notes. The same notes state that original session logs are not rewritten without a backup. That is a stronger claim than most desktop wrappers make, and the upgrade guide at docs/upgrade-and-rollback.md is where the mechanics are documented.
Skills resolution was unified in 4.0. The project scans project .dsh/skills, user DSH Skills and user Agents Skills through one directory model, so the conversation menu and the skill center show the same sources and scopes. Before that, the README implies, a skill could appear in the sidebar and be missing from the skill center.
Installing it and running a first session
Distribution is through GitHub Releases. The README points at the latest release page and notes that downloads from GitHub can be slow, in which case the project's QQ group distributes the same installers. Windows 10 and 11 x64 is the stable target; macOS Apple Silicon arm64 and Linux x64 are labelled Preview, and the README says all three must pass before they enter the same GitHub Release.
On Windows the flow is an EXE installer. There is no documented command-line install path for the desktop app itself, so the release asset is the entry point:
# Download the Windows x64 installer from the releases page
# https://github.com/ningbainb/deepseek-harness-desktop/releases/latestAfter launching, the first decision is which model to use. The model entry list has a defined ordering: models the user pinned manually come first, then the bai provider, then the remaining providers in their existing order. The settings page exposes a ChatGPT login that starts an OAuth flow through the system browser for Codex models.
The repository itself is a pnpm workspace, and the root package.json shows how the desktop package is built and packed if you want to work from source rather than from the installer:
pnpm install
pnpm desktop:dev
pnpm desktop:packThose scripts map to the @linxin666/dsh-desktop package. The root manifest also carries a set of check scripts, including runtime-deps:check, runtime-graph:check and dsh-imports:check, which the maintainers use to keep the desktop layer from drifting away from the runtime it depends on. For a first real task, the README describes the task board as the place to send work to an actual DSH Agent Session, with the run and its evidence recorded afterwards.
Browser Use and Computer Use, and their approval gates
The 4.1.0 release adds two control surfaces. Browser Use defaults to the official Playwright MCP and runs in a visible, isolated system browser session. It discovers Chrome, Edge or Chromium automatically. Clicks, typing, uploads and downloads still require approval, so the agent can look without acting until you confirm.
Computer Use defaults to a Cua Driver Native component installed with the package, with an external MCP available as an isolated fallback. Screenshots and window enumeration are observable. Mouse and keyboard actions are confirmed one at a time. If consecutive launches fail, the system drops into a safe mode rather than retrying indefinitely.
Telemetry for these events is deliberately narrow. The release notes say the new control events use a fixed result vocabulary and do not collect URLs, domains, window titles, screenshots, prompts, tool arguments, paths or credentials. Whether that is sufficient for your environment is a judgement you make against PRIVACY.md, not something the README settles.
The practical constraint is latency. Per-action confirmation is the right default for a tool that can move your mouse, but it makes long scripted sequences tedious. The README does not document a way to pre-approve a batch of actions.
Where the pinned runtime becomes your problem
The desktop host is coupled to one runtime version. The 4.1.0 notes describe the move to @deepseek-ai/dsh@0.1.6-alpha.1 as a breaking update and list the subsystems that had to be migrated. That is honest, and it also tells you what a future runtime release will cost: the desktop layer has to be adapted before you can use it.
If you depend on a different DSH version, or on plugins written against an older contract, the bundled runtime is a constraint rather than a convenience. The project's own check scripts exist precisely because the boundary between the desktop host and the runtime is a place where drift happens.
Platform support is the second limit. Windows x64 is the stable release. macOS arm64 and Linux x64 are Preview, which in this project's terms means they ship separately and are not yet promoted into the same release as Windows. If your team is mostly on Linux, this is not a drop-in replacement for a source install of the runtime.
There is also a repository-hygiene observation. The top level contains a large number of patch files, diagnostic scripts and bug-report documents alongside apps/ and docs/. That is a working repository, not a polished one, and it tells you the project moves by accumulating fixes in place. Nothing in the README says these are cleaned up before release.
Compared with running DSH from source
The alternative is the runtime's own web surface, installed and started by hand. You install Node.js and a package manager, install the DSH runtime, run it, and open the web UI in a browser. You get exactly the runtime version you chose, upgrade on your own schedule, and carry no Electron layer between you and the agent.
What you give up is everything the desktop host adds: the bundled environment, the plugin and skill lifecycle management, the task board with scheduled runs, the Git graph and SCM panel, the file preview surface covering Markdown, HTML, Diff, CSV, PDF, Office and images, the SSH panel with Web Terminal, SFTP and port forwarding, mobile remote control over a QR code, and the QQ Bot connector. The README presents the desktop app as a workbench around the runtime, not a thin launcher, and the feature list supports that reading.
A browser tab is also cheaper to keep current. The desktop app has to be rebuilt and re-released for each runtime contract change, and the release history shows that cadence: 4.0.0 and 4.1.0 landed one day apart in September 2026.
Licence, maintenance and what an upgrade actually costs
The project is BSD-3-Clause. That is a permissive licence, and for a desktop client it means you can redistribute and modify it, subject to the usual conditions about retaining the copyright notice and licence text. The repository also carries a NOTICE.md and a PRIVACY.md, and the README links the privacy policy for the telemetry described above. This is a description of what the repository states, not legal advice; if you plan to ship a modified build inside a company, read LICENSE and NOTICE.md yourself.
The last push was on 2026-09-16, and the 4.1.0 release was published the same day. The repository is not archived. That is a fast-moving project by any measure, and the practical consequence is that upgrade cost is not zero. The README points to docs/upgrade-and-rollback.md for the upgrade and rollback procedure, and the 4.1.0 notes describe backup-before-modify and rollback-on-failure for session, workspace, model, credential reference, skill, plugin, skin and pet state.
What the README does not document is a long-term support branch for older runtime versions. If you need to stay on a previous DSH contract, the upgrade guide is the document that tells you whether a downgrade path exists. Read it before you install, not after an upgrade goes wrong.
Editorial conclusion
Adopt it if you want the DSH runtime, a Web UI, plugins and skills on a Windows 10 or 11 x64 machine without assembling Node, Git and pnpm yourself, and if you accept that the desktop host tracks a pinned alpha of the runtime. Skip it if you need a vendor-supported client: the README states this is not the official DeepSeek desktop product, and it deliberately avoids the official application identity, data directories and update source. Before installing, read docs/upgrade-and-rollback.md and check whether the pinned @deepseek-ai/dsh@0.1.6-alpha.1 matches the runtime you already use, because the 4.1.0 notes describe that bump as a breaking change.
Frequently asked questions
How do I use DeepSeek Harness Desktop?
Install the Windows x64 EXE from the latest GitHub release, launch it, and pick a model from the model entry list, where manually pinned models sort first and the bai provider comes next. Work is organised through the task board, which sends tasks to a real DSH Agent Session and records the run and its evidence.
What is DeepSeek Harness Desktop and how does it work?
It is a community-maintained open source desktop client that runs the DeepSeek Harness Web Surface inside an Electron window while a desktop host manages a local DSH Runtime. The host also manages plugin installation and lifecycle, and it uses its own desktop profile so an existing DSH configuration is not overwritten.
Is DeepSeek Harness Desktop the official DeepSeek client?
No. The README states that DeepSeek has not released a standalone desktop product and that this project is not an official client. It says the project does not use or overwrite the application identity, data directory or update source of any future official desktop build.
Which platforms does DeepSeek Harness Desktop support?
Windows 10 and 11 x64 is the stable release. macOS Apple Silicon arm64 and Linux x64 are provided as Preview builds, and the README says all three platforms must pass before they enter the same GitHub Release.
Which version of the DSH runtime does DeepSeek Harness Desktop 4.1.0 use?
The 4.1.0 release notes state that the build pins @deepseek-ai/dsh@0.1.6-alpha.1 and describes that move as a breaking update, with Agent, Session, PTC, Workflow, Sandbox and Agent Team migrated to the new contracts.
Can I import existing Claude Code or Codex projects into DeepSeek Harness Desktop?
The README describes read-only discovery of projects and historical sessions, with a preview step before importing into a Harness workspace. Sensitive information is redacted during import, and historical tool calls are not re-executed.
Community notes