DeepSeek Harness: What It Is, Setup, Cost and Limits

DeepSeek Harness is DeepSeek's free, open-source AI agent. It runs on your computer and can read and edit files, run commands, search the web, schedule tasks and hand work to subagents. It works with DeepSeek's own V4.1 Flash and V4 Pro models, plus Claude, OpenAI, Kimi and GLM. You can use it in a browser tab, as a desktop app on macOS and Windows, or from the terminal with the dsh command.

The tool costs nothing, under an MIT license. You pay only for model usage, and DeepSeek's API is very cheap. A small coding task that would cost about $1.90 on Claude Sonnet 5.5 costs around $0.28 on DeepSeek V4.1 Flash at peak rates, by our math below.

The catch is maturity. As of October 30, 2026, Harness is a developer preview with no stable release, and DeepSeek's own safety notes say it "must not be treated as secure or production-ready."

How we researched this: DeepSeek's Harness page, the GitHub repository including its README, docs and SAFETY.md, DeepSeek's API pricing and privacy policy, checked October 30, 2026. We haven't run our own benchmarks; the cost figures are our calculations from list prices.

What is DeepSeek Harness?

It's an agent "harness": the software around a model that gives it tools, memory, permissions and a place to work. The model does the thinking; the harness lets it act. Claude Code and OpenAI's Codex are harnesses too, built for their own companies' models.

DeepSeek describes Harness as an open-source agent harness "with composable plugins that extend what agents can do." Its repo tagline is "Everything is a Plugin". Tools, model providers, MCP support and even the interface are plugins you can swap or write yourself.

It's broader than a coding agent. DeepSeek's page lists everyday work (files, data analysis, documents, slides), coding (fixing bugs, building features, running tests), research with citations, and scheduled background tasks.

Some numbers for context:

• The repository was created on August 13, 2026.
• It had about 266,000 GitHub stars and 32,000 forks on October 30, 2026.
• Every release so far is a pre-release. The latest is dsh-v0.2.1-rc.5, published October 30, 2026.
• The README warns: "THERE WILL BE COMPATIBILITY-BREAKING CHANGES."

What can DeepSeek Harness do?

From the repo's tool catalog, the main capabilities are:

Both DeepSeek API models have a 1M-token context window, and Harness compacts the conversation automatically as it fills up.

If MCP is new to you, our guide to what an MCP server is explains how those connections work and what to check before adding one.

How to install DeepSeek Harness

There are three ways in. All of them need a DeepSeek API key from platform.deepseek.com, or a key for another supported provider.

Web UI (quickest)

With Node.js installed, run:

npx @deepseek-ai/dsh web

It opens at http://127.0.0.1:3080 in your browser. Go to Settings, then Models, paste your API key and save. No restart is needed.

Desktop app

DeepSeek offers DeepSeek Harness desktop installers for macOS on Apple silicon and Windows 64-bit, on the Harness page. There's no Intel Mac or Linux desktop app listed; on those, use the web or terminal version.

Terminal and from source

The dsh command has several modes, including tui for a terminal interface, headless for scripts and acp for editor integrations. To build from source:

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

Before you run it on a real machine, read the next two sections.

How much does DeepSeek Harness cost?

The software is free. Model usage is billed by DeepSeek per million tokens, with off-peak hours at half price. Prices from DeepSeek's pricing page as of October 30, 2026:

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Everything else is off-peak, including weekends and Chinese public holidays. That timing matters more than it looks: for someone in the US, almost the entire working day is off-peak. In most of Europe, peak covers your morning.

Web search and fetch through DeepSeek's API cost $0.75 per 1,000 requests each. Calls made from Harness, Claude Code or Codex are free until November 1, 2026, when per-request billing starts.

What does a coding task cost?

To compare with Claude Code, we used the same small task as in our guide on whether Claude Code is free: 20 model requests, each reading 40,000 tokens and writing 1,500. That's 800,000 input and 30,000 output tokens. We priced all input as cache misses, the worst case; real sessions reuse a lot of context and cost less.

Price per task isn't price per result, though. A model that needs three attempts costs three times as much. DeepSeek doesn't publish benchmark scores for Harness itself, so judge quality on your own tasks before moving real work over.

Is DeepSeek Harness safe to use?

Treat it as experimental. DeepSeek's SAFETY.md is unusually blunt. It says Harness "has not undergone a security audit and must not be treated as secure or production-ready." It can run model-generated commands, load third-party plugins, and reach your network, files and credentials. The same file says sandboxing and approval prompts "do not guarantee isolation or prevent damage."

DeepSeek recommends least privilege, a disposable VM or container, and backups. We'd add:

1. Don't give it danger-full-access on your main computer.
2. Install plugins only from sources you trust, since plugin installs can run build scripts.
3. Keep API keys for production systems out of the environment it runs in.

The same advice applies to any agent with no approval step. Our guide to Claude Code's dangerously skip permissions flag covers how to sandbox an agent properly.

Where does your data go?

Harness itself runs locally, and the web UI listens only on your own machine. Your prompts and code go to whichever model provider you configure.

If that's DeepSeek, its privacy policy says it collects, processes and stores personal data in the People's Republic of China. The policy also says data may be used to "train and improve our technology", with a right to opt out. If that's a problem for your company or your clients' code, point Harness at Anthropic or OpenAI instead, or don't use it for that work.

DeepSeek Harness vs Claude Code and Codex

Choose DeepSeek Harness if you want the cheapest capable models, the freedom to swap providers, or a plugin system you can change yourself. Choose Claude Code or Codex if you need a stable tool for client or company work, or if the data-location question rules DeepSeek out. Our Claude Code vs Codex comparison covers those two in depth, and OpenCode vs Claude Code covers another open-source option.

FAQ

Is DeepSeek Harness free?

Yes. The software is free and open source under the MIT license. You pay for model usage through the DeepSeek API or whichever provider you connect, from $0.15 per million input tokens off-peak on V4.1 Flash.

Is DeepSeek Harness open source?

Yes. The full source is on GitHub under the MIT license, at github.com/deepseek-ai/deepseek-harness.

Does DeepSeek Harness work with Claude or local models?

It has built-in providers for Anthropic, OpenAI, Kimi and GLM, so you can run it with Claude models. Local models aren't listed as a built-in provider, but you can add any server that speaks the OpenAI or Anthropic API format as a custom provider.

Is there a DeepSeek Harness desktop app?

Yes, for macOS on Apple silicon and Windows 64-bit, downloadable from DeepSeek's Harness page. Linux and Intel Mac users can run the web UI or the dsh command instead.