DeepSeek Harness Guide (2026): Install, Models, Pricing, Privacy and vs Claude Code

DeepSeek Harness is the open-source agent app from DeepSeek, and it is growing fast. DeepSeek says it reached 1.54 million daily active users by November 4, 2026, burning about 92 trillion tokens a day.

In this guide I explain what Harness is, how to install it, which models it uses, what it costs, and how it compares with Claude Code and Codex.

What DeepSeek Harness Is

DeepSeek Harness (the command is dsh) is an agent harness. It is not a model. It is the runtime that connects a model to your files, terminal, tools and session history so it can do multi-step tasks.

Its slogan is “Everything is a Plugin”. It is built on the Cordis plugin framework, so the agent loop, model adapters, tools, storage and even the UI are all swappable plugins.

• License: MIT, with the code on GitHub at deepseek-ai/deepseek-harness.
• Interfaces: a desktop app, a browser Web UI, a headless mode and an SDK.
• Jobs it does: coding, file work, web search, scheduled tasks and office document previews.
• Status: developer preview. The repo warns that compatibility-breaking changes will happen.

The Growth Numbers and Timeline

Harness went from developer toy to mass product in under three months. The desktop app was the turning point, because it removed the need for Node.js and a terminal.

DeepSeek also says users have built about 35,000 plugins, and about 60% of active users have created or installed at least one. An official plugin marketplace and creator program are being prepared, but no launch date has been given.

Note: these figures come from DeepSeek and were reported by Bloomberg and others. They have not been independently audited, and the token count covers official DeepSeek API usage only.

How to Install DeepSeek Harness

You have two routes. The desktop app is easiest. The npm route suits developers, servers and Linux users.

Option 1: the desktop app (Windows and Mac)

1. Download from an official source. Use the product page at deepseek.com/en/harness or the GitHub releases page. Third-party copies are circulating, so avoid them.
2. Check your system. The desktop build supports Windows and Apple-silicon Macs on macOS 13 or later. It bundles its own Node runtime.
3. Sign in. Log in with a DeepSeek account (it shows your credit) or paste a DeepSeek API key.
4. Create a workspace. Point it at a dedicated folder, not your whole home directory.
5. Start a session. Pick a model and agent preset, then type your task.
6. Optional CLI. Since v0.2 rc.2, the desktop menu can install the dsh command for your terminal.

Option 2: npm and the browser Web UI

1. Install Node.js. Recent releases need Node.js 24.9 or newer (the 0.2.2 alpha raised the minimum). Check with node --version.
2. Run it once with npx, or install it globally (commands below).
3. Set your key. Export DEEPSEEK_API_KEY, or add it later in Settings → Models.
4. Open the Web UI. dsh web serves it at http://127.0.0.1:3080 and opens your browser.
5. Install pnpm if you plan to add or remove plugins, because Harness uses it for that.

# quick start, no install
npx @deepseek-ai/dsh web

# or install globally
npm install -g @deepseek-ai/dsh
npm install -g pnpm
dsh --version

# set your key (macOS / Linux)
export DEEPSEEK_API_KEY="your_key_here"
dsh web

On Windows PowerShell, set the key with $env:DEEPSEEK_API_KEY="your_key_here" instead. Add --no-open if you do not want a browser tab to open, for example over SSH.

Tip: New to DeepSeek keys? My DeepSeek API Python tutorial shows how to create one and add credit.

Models and Providers You Can Use

Harness ships with DeepSeek-V4.1-Flash as the default model. On the API this model is called deepseek-flash. The stronger deepseek-v4-pro is also available.

You are not locked to DeepSeek. Open Settings → Models to add other providers with your own keys:

• Built-in providers: DeepSeek, Anthropic, OpenAI, Moonshot (Kimi) and Zhipu (GLM).
• Custom endpoints: anything that speaks OpenAI Chat Completions, OpenAI Responses or Anthropic Messages.
• Local models: point a custom endpoint at a local OpenAI-compatible server.
• Not yet: OAuth sign-in providers such as a ChatGPT or Codex subscription.

Model changes apply to the next request without a restart.

Plugins, Presets and Key Features

Because everything is a plugin, the feature list depends on what you install. These are the main built-in and official parts.

• Agent presets: Standard, PTC, Minimal and Creator. Each one controls which tools a session can use.
• Official plugins: Shell, Agent loop, Subagent, Web search, Agent Teams (beta) and Voice input (beta, transcribed locally).
• Creator mode: an experimental mode where the agent writes or edits plugins for you.
• Scheduled tasks: recurring jobs as often as once a minute, off by default.
• Background runs: tasks keep running when you close the window.
• Trajectory view: an append-only log of every model turn and tool call, with resume, fork and replay.
• Office previews: Word, Excel, PDF, HTML, Markdown and code files.

What is new in the last few weeks

The 0.2.2-alpha.2 release on October 28 is the latest notable update. It is still an alpha, so treat these as experiments.

• External CLI delegation: a subagent_oar plugin can hand tasks to a local Codex CLI or Claude CLI.
• Worktrees: an experimental Change Request and Git worktree flow.
• Developer Statistics: token, cache and latency stats per turn.
• Breaking changes: Node.js 24.9 minimum, and old tools.mode settings must be migrated.

To add a community plugin to the Web UI profile, the pattern is dsh plugin --profile web add <package>. Plugin authors tag their GitHub repos with the dsh-plugin topic.

Pricing: What You Actually Pay

The Harness software is free. Your cost is the model tokens you use, billed by whichever provider you connect. DeepSeek has no separate Harness subscription.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Every other hour bills at the half price rate. Full details are in my DeepSeek API pricing guide.

Warning: agents use far more tokens than chat. DeepSeek’s own numbers work out to roughly 60 million tokens per active user per day, so set a spending limit and watch your balance.

Privacy and Safety

Harness stores sessions, tool records and settings locally by default. You can turn off its anonymous reports about settings and project lists.

That does not make it private end to end. Whatever model provider you connect sees your prompts and the code you send. Web tools, MCP servers and plugins can also send data out under their own policies.

• No security audit: DeepSeek says Harness must not be treated as secure or production-ready.
• Real file access: the agent runs model-written commands and can change or delete files in its workspace.
• Plugin risk: third-party plugins run with real access, and quality varies a lot.
• DeepSeek servers: using DeepSeek’s API sends data to DeepSeek, which some companies do not allow.

DeepSeek’s advice is least-privilege use, ideally in a disposable VM or container, with backups of anything the agent can reach. Keep production credentials out of the environment.

DeepSeek Harness vs Claude Code vs Codex

All three are agent tools that read your repo and run commands. The big differences are openness, default model and cost.

Pick Harness if you want the cheapest tokens and full control over the runtime. Pick Claude Code or Codex if you want a polished, supported tool and stronger frontier models.

My Take

Harness matters because it turns the agent itself into open, hackable software. Claude Code and Codex treat the loop as a product. DeepSeek treats it as a plugin you can rewrite.

The cheap Flash model plus a free app is a strong combo for hobby projects and bulk tasks. For anything sensitive, I would wait for a security audit and a stable 1.0. The breaking changes alone make it a moving target today.

Troubleshooting: Common Problems and Fixes

• npx fails or plugins will not install: update Node.js to 24.9 or newer and install pnpm with npm install -g pnpm.
• Web UI does not open: browse to http://127.0.0.1:3080 by hand. Over SSH, forward that port and use --no-open.
• “Model not found” after an update: some old model IDs were removed. Reselect a model in the session picker.
• Agent stops mid-task: type “continue”, or split the task into smaller steps.
• Insufficient balance errors: top up your DeepSeek account. My DeepSeek API errors guide covers 402 and 429 codes.
• Old settings stopped working: migrate tools.mode and DSH_TOOLS_MODE after the 0.2.2 changes.
• Mac app will not run: the desktop build is Apple-silicon only. Intel Macs should use the npm Web UI.

Frequently Asked Questions

Is DeepSeek Harness free?

Yes, the app is free and MIT licensed. You only pay for the model tokens you use with DeepSeek or another provider.

Does DeepSeek Harness work on Linux?

The desktop app targets Windows and Apple-silicon Macs. On Linux you can run the browser Web UI with npx @deepseek-ai/dsh web and Node.js.

Which model does Harness use by default?

It defaults to DeepSeek-V4.1-Flash. You can switch to deepseek-v4-pro or add Anthropic, OpenAI, Kimi, GLM or a custom endpoint.

Can I use Claude or GPT models in Harness?

Yes. Add Anthropic or OpenAI in Settings → Models with your own API key. Subscription OAuth sign-in is not supported yet.

Is DeepSeek Harness safe for work code?

DeepSeek says it has not had a security audit and is not production-ready. Use a separate workspace or VM and keep secrets out of reach.

How many people use DeepSeek Harness?

DeepSeek reported 1.54 million daily active users and about 92 trillion tokens per day as of November 4, 2026. These figures are not independently audited.

Is there a Harness plugin marketplace?

Not yet. DeepSeek says an official marketplace and creator program are in preparation, with no launch date announced.

Sources

• GitHub: deepseek-ai/deepseek-harness
• DeepSeek API Docs: Models and Pricing
• DataCamp: What Is DeepSeek Harness?
• DataCamp: DeepSeek Harness tutorial
• NYU Shanghai RITS: Harness v0.2 desktop app
• KuCoin News (citing Bloomberg): 1.54M DAU, 92T tokens
• BestHub: DeepSeek Harness 0.2.2 release notes