DeepSeek Harness is an open source agent harness from DeepSeek: a local runtime that turns a language model into an agent that can read your repo, run commands, and keep a plan. You install it as a desktop app or launch a web UI from your machine. It is in public preview, and DeepSeek ships it as open source on Cordis, with an "everything is a plugin" design.
That is the short version. The rest of this guide is for developers who searched "deepseek harness" because they want to try it, not because they want another everyday productivity pitch. You will see how to install it, what still needs a DeepSeek account or API key, what typically leaves your machine, and where it sits next to Claude Code style agents.
What DeepSeek Harness actually is
A harness is the layer around a model that makes tool use real. The model proposes steps. The harness reads files, runs shell commands, tracks sessions, and shows you diffs. DeepSeek Harness (often shortened to dsh) is DeepSeek's open source version of that layer.
On the product page, DeepSeek lists five job areas the preview aims at:
- Everyday work with docs, data, and slides
- Coding: explore repos, fix bugs, build features, run tests
- Research with cited sources
- Background tasks and batch scripts
- Plugins you install or build yourself
The coding path is why most developers land here. The same product page also markets office workflows. Treat those as optional, not as the reason you install it.
DeepSeek built the runtime on Cordis and markets the architecture as "everything is a plugin." In practice that means the model adapter, the tool registry, and even the agent loop are swappable plugins, not a sealed proprietary shell.
How to install DeepSeek Harness
You have two supported starts, both from DeepSeek's own pages.
Desktop app (preview). Download the installer from deepseek.com/en/harness. Current coverage called out in secondary reporting for the v0.2 wave is macOS on Apple silicon and Windows on x64. The desktop build bundles the dsh command, so you do not need a separate Node.js install for that path. Check the download page again before you ship internal docs, because preview installers move.
Web UI from npm. Install a current Node.js, then run:
npx @deepseek-ai/dsh webDeepSeek's GitHub README says this starts the Web UI at http://127.0.0.1:3080 by default and opens your browser on a local launch. Pass --no-open if you only want the server.
From a source checkout, the same README documents the longer path: clone deepseek-ai/deepseek-harness, pnpm install, pnpm run build, then pnpm dsh web.
DeepSeek labels the whole thing a public preview. Expect compatibility breaking changes between builds. Do not pin production workflows to a single preview build without a rollback plan.
What you still need: account or API key
The harness is local. The model usually is not.
DeepSeek's own UI copy and provider docs assume you either:
- Sign in with a DeepSeek account, or
- Paste a DeepSeek API key under Settings → Models (default endpoint
https://api.deepseek.com)
You can also add third party providers or a custom OpenAI compatible endpoint. The provider guide covers Anthropic, OpenAI, and other catalog ids, plus a custom base URL when you run your own gateway.
So "local" means the agent loop, file access, and UI run on your machine. Inference still goes to whichever provider URL you configured, unless you point that URL at a model you host yourself.
DeepSeek's Harness marketing page does not publish a public Harness specific fee schedule in the copy we checked. Account versus API billing follows whatever DeepSeek (or your other provider) lists in its own billing docs on the day you sign up.
What data leaves your machine
This is the question developers ask after the install screenshot.
Stays local by design: the workspace files the agent reads, the shell the agent runs, session logs and traces you inspect in Developer tools, and plugins installed into the local runtime.
Leaves when you use a remote model: prompts, tool results, and file excerpts the agent sends to the model endpoint. If you use DeepSeek's hosted models, that traffic goes to DeepSeek's API. If you point the harness at another provider, the same payloads go there instead.
Gray area you should verify yourself: web search or other network tools enabled by your account or plugins. DeepSeek's v0.2 rc.1 release notes say account models can get web search without an extra key in some preview builds. Treat that as a network egress path, and confirm it in your Settings before you open a private repo.
Practical rule: open a throwaway repo first. Watch the model provider traffic. Then decide whether the harness meets your company's data policy.
Plugins, Creator mode, and developer tools
DeepSeek's pitch is that you extend the harness the same way you extend an editor: install a plugin, or ask the agent to write one.
Creator mode lets you describe a plugin in chat. The agent loads plugin development skills, writes the package, installs it, and verifies it, as shown on the product page with a Pomodoro timer example.
Official and experimental plugins on the marketing page include scheduled tasks, voice input, terminal helpers, agent teams, and auto approval review. Scheduled tasks are the clearest "leave it running" feature: you set a recurring prompt and inspect run history later.
Developer tools expose execution traces, tool call payloads, timing, and hierarchy. If you are comparing harnesses, this is the part that matters more than the slide generator. You can see what the model saw, what tool ran, and how long each step took.
Secondary coverage of the v0.2.1 alpha also mentions an experimental Claude Code Mods compatibility layer. DeepSeek frames it as a test, not a promise of full compatibility. Keep that label when you write internal notes.
When DeepSeek Harness is a fit (and when it is not)
Fit: you want an open source agent loop you can run locally, inspect, and extend with plugins. You are fine authenticating to DeepSeek or wiring your own model endpoint. You want coding and repo work first, with office plugins as extras.
Not a fit yet: you need a generally available, contract locked coding agent with a stable plugin API. You cannot send any repo context to a remote model. You only want a chat box in the browser with no local runtime.
If you are mapping the wider coding agent landscape, start with our roundup of the best AI coding agents, then come back here for the DeepSeek specific install and data questions.
Next step if you want to build this kind of system
If the install worked and you now want the engineering path behind agents, RAG, and production workflows, explore AI Engineering or the self paced AI Flex catalog. Both are built for people who already write code and want agent systems, not a no code starter track.
Go deeper
- Best AI coding agents
- What is an MCP server
- AI agent memory: how it works
- How to become an AI engineer if you want the career path behind tools like this