DeepSeek Harness: How to Install and Use It on Windows, Mac and Linux (2026 Guide)

DeepSeek Harness is DeepSeek’s free, open-source AI agent. It can write code, work through files, research topics and run scheduled tasks on your computer. On 29 September 2026 the v0.2 preview added proper desktop apps for Windows and macOS, so you no longer need a terminal to try it. This guide explains how to install DeepSeek Harness on the desktop or with npx, connect a model, add plugins, and keep costs and risks under control.

Facts in this guide were checked on 2 October 2026 against the official DeepSeek Harness GitHub repository, the DeepSeek Harness download page and DeepSeek’s API pricing page. Where a detail comes from an independent guide instead, the text says so. Version 0.2 is a preview, so menus may change.

What Is DeepSeek Harness?

An AI model on its own only produces text. A harness is the software around the model that lets it act: reading and editing files, running commands, searching the web, remembering a session and asking you for permission. DeepSeek Harness (its command is dsh) is DeepSeek’s own harness, released as open source under the MIT licence.

Its defining idea is what the README calls an “everything-is-a-plugin” architecture, built on a framework called Cordis. Models, tools, skills and even parts of the interface are plugins that can be swapped. That is why the community has built so many add-ons so quickly. Developers tag them with the dsh-plugin topic on GitHub.

DeepSeek says Harness is now the most-used coding agent among people who use its API, measured by daily active users and sessions. It also says about 60% of users run third-party plugins. Those are company figures, not independent measurements.

What changed in v0.2

• Desktop installers for macOS (Apple Silicon, .dmg) and Windows (64-bit, .exe).
• Plugin management: install, disable and remove plugins by typing their npm package name.
• Automation tasks: a built-in plugin for scheduled and recurring jobs, with execution logs.
• Office-style work: upload documents, spreadsheets or PDFs and ask it to organise material, analyse data, create charts or draft presentations.
• Better file display and previews inside the app.

For context on why agents like this matter, see our explainer on how AI agents are changing business.

Is DeepSeek Harness Free?

The software is free. It is MIT-licensed, and there is no subscription for the harness itself. You pay only for the AI model it uses. You have three options:

• DeepSeek’s API: the default, paid per token.
• Another provider: independent guides show it working with OpenRouter and other OpenAI-compatible services.
• A local model through Ollama, LM Studio, vLLM or llama.cpp: no API bill, but you need capable hardware.

DeepSeek’s official pricing page listed these rates per million tokens when we checked:

DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. The morning window (06:00–10:00 UTC) is 07:00–11:00 in the UK and Ireland and 08:00–12:00 in Germany, France, the Netherlands, Spain, Italy, Sweden and Poland during summer time. If you are in Europe, running long agent jobs in the afternoon is the simplest way to halve the cost. Agents use far more tokens than a normal chat, because every step re-sends context, so this matters more than it looks.

How to Install DeepSeek Harness

There are three ways to install it. Pick the one that matches how comfortable you are with a terminal.

Option 1: Desktop app (Windows and Mac, easiest)

1. Go to the official page at deepseek.com/harness. Download only from there or the official GitHub repository; unofficial “DeepSeek desktop” builds also circulate on app-listing sites.
2. Choose the Windows 64-bit .exe or the macOS Apple Silicon .dmg.
3. Run the installer and open the app.
4. Add your model and API key (see “First run” below).

At launch there is no official Linux desktop installer and no Intel Mac build. On Linux, use option 2.

Option 2: npx (Windows, Mac and Linux)

If you have Node.js installed, one command starts DeepSeek Harness in your browser:

npx @deepseek-ai/dsh web

The README says the web interface opens at http://127.0.0.1:3080 by default. To pin a version, use npx @deepseek-ai/dsh@<version>.

Use a current Node.js release. An independent ComputingForGeeks guide says it needs Node 22 or Node 24 and later (not Node 23), and reports that the older Node 18 shipped with some Linux distributions fails at launch.

Option 3: Build from source (developers)

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Building from source is useful if you want to write plugins or follow the latest changes.

How to Use DeepSeek Harness: First Run

1. Connect a model

Open Settings → Models and paste an API key. For DeepSeek’s own models, create the key at platform.deepseek.com and add some credit (our DeepSeek API guide walks through keys and pricing in detail). Start with deepseek-flash: it is far cheaper, and you can switch to v4-pro for harder tasks later.

2. Choose a workspace folder

Point DeepSeek Harness at one project folder, not your whole home directory. The agent can read and edit files there, so a dedicated folder limits what it can touch.

3. Pick a mode

Independent guides describe several runtime modes:

• Standard: the full toolset for coding and general tasks.
• Code: focused on programming work.
• Minimal: just a shell and an editor, for lightweight use or small local models.
• Creator: for building your own plugins, including by asking the agent to write them.

4. Give it a clear, bounded task

Good first prompts are specific and checkable:

• “Read this repository and explain its structure in five bullet points.”
• “Add input validation to signup.js and write tests for it. Show me the diff before saving.”
• “Analyse sales.xlsx, summarise the monthly trend and draw a chart.”

Read every change before accepting it. A harness is only as safe as the permissions you give it.

5. Schedule a recurring task

With the v0.2 automation plugin you can turn a prompt into a scheduled job, such as a weekly dependency check or a daily report from a folder of CSV files. You can then review the run logs and adjust how often it runs.

DeepSeek Harness Plugins

Plugins are where DeepSeek Harness gets interesting. Because almost everything is a plugin, you can add web search, browser control, MCP servers, new model providers, interface themes or chat integrations.

How to install a plugin

1. Find a plugin on the dsh-plugin GitHub topic or in the official plugin list.
2. Note its npm package name.
3. In the desktop app, open plugin management and enter that name. You can disable or uninstall it the same way.
4. Some plugins need their own key; web-search plugins, for example, usually need a search API key.

Plugin safety

A plugin is code that runs with your agent’s permissions. Before installing one:

• Prefer official plugins or ones with an active repository, recent commits and real users.
• Read what the plugin can access.
• Avoid plugins that ask for broad system access without a clear reason.

DeepSeek itself lists better sandboxing and security among its next priorities, which tells you the current version is still early. If you want an extra layer of isolation for agents, our guide to NVIDIA OpenShell, an open-source agent sandbox, explains one approach.

Using DeepSeek Harness With a Local Model (Ollama)

Running a local model means no API costs and your code never leaves your machine. New to local models? Our guide to running an LLM locally covers installing Ollama and choosing a model size. According to the ComputingForGeeks walkthrough:

1. Settings live in ~/.dsh/settings.yaml.
2. Add a provider entry for Ollama with its baseURL (normally http://127.0.0.1:11434/v1) and at least one model.
3. Set the model ID exactly as Ollama reports it, for example qwen3:8b rather than qwen3.
4. Set an API-key environment variable even though Ollama ignores it (for example OLLAMA_API_KEY=ollama); otherwise the first run fails.
5. Give the model a large enough context window, because agents need a lot of context.

Be realistic about quality. Small local models struggle with long, multi-step agent tasks. They are fine for experiments and simple edits; serious coding work usually goes better with a strong hosted model.

DeepSeek Harness vs Claude Code and Other Agents

“DeepSeek Harness vs Claude Code” is one of the most common searches about the tool. The honest comparison is about trade-offs, not a single winner:

Choose DeepSeek Harness if low running costs, open-source code, model choice or local models matter most to you. Choose Claude Code if you already pay for Claude and value Anthropic’s models; our Claude pricing guide breaks down what that costs. Many developers try both on the same task before deciding. If you have not tried Anthropic’s tool yet, our Claude Code tutorial covers setup and the essential commands.

Common Problems and Fixes

• “API key is invalid”: check the key has no spaces, belongs to the provider you selected and the account has credit.
• Fails at launch with npx: upgrade Node.js; old versions are a common cause.
• “Failed to load plugins” or “unknown tool”: disable recently added plugins one at a time, then update DeepSeek Harness.
• Local model never uses tools: your model server may need tool-calling enabled. The vLLM server, for example, needs its tool-choice flags turned on.
• Output stops halfway: you may have hit the output token limit; ask it to continue or split the task.

Who Should Use DeepSeek Harness?

It is a strong fit for developers who want an open-source coding agent, cost-conscious teams, people who want to run agents against local models, and tinkerers who enjoy building plugins. With the new desktop app, non-developers can also use it for document and spreadsheet work.

Be more cautious if you handle sensitive company data. As with any AI service, check where your prompts and files are processed and what your organisation’s policies allow, especially under European data rules.

If you are building a wider AI toolkit, our roundup of the best AI tools for work and productivity covers the other categories. Harness can also call creative tools through plugins, such as image workflows from our ComfyUI beginner’s guide.

FAQ

Is DeepSeek Harness free?

Yes, the software is free and MIT-licensed. You pay for the model you connect, usually per token through DeepSeek’s API, unless you run a local model.

Is DeepSeek Harness open source?

Yes. The full source code is on GitHub at deepseek-ai/deepseek-harness under the MIT licence.

Does DeepSeek Harness work on Windows and Mac?

Yes. Version 0.2 offers a Windows 64-bit installer and a macOS installer for Apple Silicon. On Linux, or on an Intel Mac, run it with npx @deepseek-ai/dsh web.

Can DeepSeek Harness use models other than DeepSeek?

Yes. Its plugin system supports other providers and local models such as those served by Ollama or LM Studio.

Is DeepSeek Harness safe to use?

It is an early preview that can edit files and run commands. Limit it to a project folder, review its changes, install only trusted plugins and keep sensitive data out of its workspace.

Final Thoughts

DeepSeek Harness has gone from a developer-only command-line project to a desktop app anyone can install, without losing what made it popular: open source, cheap to run and extensible through plugins. Start with the desktop app or npx, connect deepseek-flash, give it one folder and a clear task, and add plugins only when you know what you need.

Sources: DeepSeek Harness on GitHub; DeepSeek Harness download page; DeepSeek API pricing; SMM report on the v0.2 release; ComputingForGeeks local-model guide. Checked 2 October 2026.