DeepSeek DSH — DeepSeek Harness
DeepSeek DSH, DeepSeekDSH and dsh are common search terms for DeepSeek Harness: DeepSeek AI's open-source agent harness. This page explains the project itself before sending you into installation or reference docs.
On this page
What is DeepSeek DSH?
DeepSeek DSH usually means DeepSeek Harness, whose npm package is @deepseek-ai/dsh. It is not a new DeepSeek model. It is an open-source agent harness: the layer that connects an LLM to a workspace, tools, plugins, persistent sessions, permissions and an interface so the model can do multi-step work instead of only returning text.
A useful shorthand is: the model reasons; DSH provides the operating environment around that model. When you ask DSH to inspect a repository, edit files, run commands or delegate work, the harness decides which capabilities are available to the session and records the interaction around them.
DeepSeek Harness is still moving quickly. The project is in developer preview, so alpha builds can change APIs, defaults and persistence behavior. That is why version-sensitive pages on this site show the exact upstream source snapshot they were checked against.
What does DSH add on top of a model?
Calling a model API gives you model input and output. A harness adds the machinery needed to turn that model into a working agent. In DSH, the important pieces are easier to understand if you separate them by job:
Workspace + tools
The agent can work with files, shell commands, search and other mounted tools instead of being limited to a chat response.
Agent presets
A session starts with an agent composition such as Standard, PTC, Minimal or Creator. The preset controls the model-facing tool surface.
Plugins and extension points
DSH follows an “everything is a plugin” architecture, so capabilities can be composed rather than permanently hard-coded into one monolithic agent.
Sessions and control
Longer-running work needs session state, logs, reconnect behavior and a UI where the user can inspect what the agent is doing.
This is also why “DeepSeek Harness” and “DeepSeek model” should not be treated as synonyms. A model can be one component inside the harness, while the harness owns the surrounding agent workflow.
DeepSeek DSH is not the older Linux dsh command
The short name dsh is ambiguous on the wider web. Long before DeepSeek Harness, Linux and Unix users used projects named dsh for “distributed shell” style command execution across multiple machines. Those tools are unrelated to DeepSeek AI.
If the page you find talks about running the same shell command across a cluster of hosts, Debian packages such as dsh, or distributed shell administration, it is probably the older Linux meaning. If it talks about @deepseek-ai/dsh, agent presets, plugins, the Web UI or DeepSeek Harness, it is the DeepSeek project discussed here.
DeepSeekDSH is the name of this independent guide site. DeepSeek Harness is the upstream project. We are not affiliated with DeepSeek; the official repository and npm package are linked below.
How DeepSeek Harness fits together in one session
You do not need to learn the entire Cordis architecture before using DSH. For a normal session, this mental model is enough:
model
+ agent preset (tool composition)
+ workspace
+ host services / permissions
+ installed plugins and providers
= one DSH agent sessionThe preset is especially important because it changes what the model sees as available capabilities. It is not the same thing as the model selection itself, and it is not the same thing as the host permission system. This separation becomes useful once you start comparing Standard, PTC, Minimal and Creator mode.
Sessions already running on a preset keep that composition. If you duplicate or change a preset, start a new session to test the changed tool surface rather than assuming an existing session will recompose itself in place.
DeepSeek Harness currently ships four built-in agent modes
The current Web UI exposes four built-in presets. They are not four different LLMs; they are four different agent compositions.
Standard
The normal full coding-agent preset with editing, shell, file/web search, skills, planning, goals, subagents and workflows.
PTC
A full coding-agent preset where most tools are exposed through the PTC SDK so the model can combine multi-step operations inside TypeScript run_code.
Minimal
A deliberately narrow two-tool composition: persistent bash plus str_replace_editor.
Creator
Standard capabilities plus runtime inspection, plugin experiments and preset-authoring guidance for people creating custom agent compositions.
For most users, Standard is the right starting point. Switch because a workflow has a concrete requirement, not because another mode sounds more advanced.
Ways to run DeepSeek DSH
1. Desktop-first path for Windows and macOS
If you mainly want to try the product rather than manage a Node environment, start with the community-maintained DSH Desktop path. It packages the local runtime flow into a desktop app. It is independent community software, not an official DeepSeek desktop release.
Open the DSH Desktop download guide →
2. Official npm / npx path
The upstream CLI package is @deepseek-ai/dsh. With Node.js installed, the common Web UI command is:
npx @deepseek-ai/dsh webThe local Web UI uses http://127.0.0.1:3080 by default. If you prefer a full operating-system walkthrough, use the maintained installation guide rather than troubleshooting the command from memory.
Open the installation tutorial →
3. Run from source
Source checkout is mainly for contributors, plugin authors or users who need unreleased code. It adds repository and package-manager complexity, so it should not be the default route for someone who only wants to evaluate DSH.
Which DeepSeek DSH version is current?
There are two version lines worth separating. The default npm publication baseline on this site is 0.1.1-rc.2, while the latest source/prerelease snapshot we have verified is 0.1.2-alpha.5 at commit 49a606b.
That distinction matters because an alpha release can contain fixes and new behavior that the npm latest baseline does not yet represent. Do not describe “the latest DSH” without saying whether you mean the default npm tag or the faster alpha/source line.
Official DeepSeek DSH links
DeepSeek DSH FAQ
Is DeepSeek DSH a model?
No. DeepSeek Harness is an agent harness around a model. The harness supplies the session, tools, presets, plugins and operating workflow; model selection is a separate layer.
Is DeepSeek DSH the same thing as DeepSeekDSH.com?
No. DeepSeekDSH.com is this independent community documentation site. The upstream project is maintained in the deepseek-ai/deepseek-harness GitHub repository.
Is dsh the Linux distributed shell?
Not in this context. The name is shared with older distributed-shell utilities, but DeepSeek Harness is a separate AI-agent project published as @deepseek-ai/dsh.
Should a beginner use Standard or PTC mode?
Start with Standard. PTC is useful when you understand why a task benefits from code-orchestrated multi-step tool calls; it is not a universal “faster mode”.
Do I need the alpha release?
Not automatically. Alpha builds can contain important fixes, but they are developer-preview releases. Check the release note for the exact problem you need to solve before changing version lines.
Where should I go after this page?
If you are installing, go to Tutorials. If DSH already runs and you are choosing an agent setup, read the four-mode decision guide. If you have a literal error message, use the Error Center instead of a general explainer.