How to use DeepSeek Harness: setup, first build, and what it costs

The short answer: DeepSeek Harness is a free, MIT-licensed app from DeepSeek that turns an AI model into a worker that can read your project files, change them, run commands and check its own work. You install it on your own computer and add a model key. The app costs nothing; the only bill is what the model charges. It is still a developer preview, so it is for people who are comfortable with a developer's tools, not a plug-and-play product for the front office.

Checked on 11 November 2026 against DeepSeek's own pages: the project repository, the Harness page, the documentation and the API price list. Anything we could not confirm on those pages is marked as unconfirmed.

What it is, in plain words

An AI model on its own is a very capable writer that cannot touch your files. A harness is the app wrapped around the model that gives it hands: it can open files, edit them, run a command and look at the result. DeepSeek Harness is DeepSeek's own version, and the company describes it as built on an "everything is a plugin" design, meaning almost every capability can be swapped or extended.

It can write code, but DeepSeek's page also says you can use it to work with documents, analyze spreadsheets and schedule tasks. Most people who use it today are developers, because the setup and the safety decisions below assume you know what a project folder and a terminal are.

Who it suits: a developer or a technically comfortable owner who wants a free, open tool and is happy to bring their own model. Who it does not suit: anyone who needs a supported product with a help desk and a stable feature set. DeepSeek's own repository says it is in developer preview and that "there will be compatibility-breaking changes".

Step 1: install it

There are two routes, and you only need one.

The desktop app. DeepSeek's Harness page offers a download for macOS on Apple Silicon. We could confirm that one download link and no others on the vendor's page, so if you use Windows or Linux, use the web route below.

The web route. Install Node.js (a free tool that runs this kind of program), open a terminal and run one line:

npx @deepseek-ai/dsh web

That starts Harness on your own computer and opens it in your browser at an address that begins with 127.0.0.1, which means "this machine only". If you are connecting to a remote machine, add the option --no-open and open the address yourself.

Step 2: add a model

Harness needs a model to do the thinking. Open Settings, then Models, paste a DeepSeek API key and save. An API key is a long password that lets the app use DeepSeek's service and bills your DeepSeek account. You create one on DeepSeek's developer platform.

Two details from the documentation matter:

• The key is write-only. After you save it, the page shows only a hidden version, and the real key is kept in a file on your computer.
• The change takes effect on the next request, with no restart.

You are not locked to DeepSeek. The same screen has Add model provider, where you can pick from providers Harness ships with (the guide names Anthropic, OpenAI, Kimi and GLM as examples) or add a custom model API for any server that speaks the common OpenAI or Anthropic formats. That includes a company gateway or a server you run yourself.

Step 3: pick a workspace

Click Choose workspace and select the folder for the project. Harness will not let you start a session until you do. Think of the workspace as the room you are letting the assistant work in. Pick the project folder, never your whole documents folder.

Step 4: choose how much freedom it gets

This is the most important decision in the guide. DeepSeek's documentation lists three sandbox modes:

The name of the third one is the vendor's own warning. Our advice: start in read-only to let it explore and explain, move to workspace-write when you want it to build, and do not use full access on a computer that holds anything you cannot lose. On top of the mode, the app asks for your approval before actions that need it.

Be honest with yourself about the limits of these controls. DeepSeek's safety notice says Harness has not had a security audit, that it can run code and commands the model writes, and that sandboxes and approval prompts "can reduce risk" but do not guarantee isolation. Its advice is to use the least access you can, prefer a disposable virtual machine or container, keep backups, and review plugins and proposed commands before you let them run.

Step 5: use plan mode for anything bigger than a small fix

Plan mode asks the assistant to lay out its approach before it acts. One caution from the documentation: plan mode is "soft guidance". It nudges the model; it does not physically stop it from acting. The hard limits are the sandbox mode and the approval settings above, which are separate.

Step 6: a safe first build

Do your first run on a copy of a small, low-stakes project.

1. Start a session and ask it to explain the project. DeepSeek's own example is: "Summarize this repository and identify its main packages."
2. Read the answer. If it is wrong about what your project does, stop and fix the instructions before it changes anything.
3. Then ask for one small change, and ask it to write a test for that change and run the tests. A test is a small automatic check that the change behaves as intended. Asking for tests is the cheapest way to catch a confident mistake.
4. Look at the files it changed before you keep them. If your project uses version control, commit first so you can undo everything in one step.

Step 7: check its work with the trajectory view

Harness has a trajectory view. DeepSeek describes it as a way to "inspect execution traces and detailed runtime information to troubleshoot issues with tool calls and task execution". In plain terms, it is a timeline of every step the assistant took: what it opened, what it ran and what came back. When a result looks too good, read the trajectory. It is where you notice that a test was skipped or a command failed and the assistant carried on anyway.

Step 8: subagents for bigger changes

Larger jobs can be split. DeepSeek lists subagents among Harness's capabilities, and its guide says the main agent can read and edit files, run commands, delegate work and keep a plan. A subagent is a helper the main assistant hands part of the job to, so one piece does not clutter the whole conversation. Use it for jobs with clearly separate parts, such as "update the booking page" and "update the confirmation email", and review each part separately.

Step 9: plugins, skills and Creator mode

Plugins are how Harness grows. The Harness page shows official plugins, several of them labelled Experimental: agent teams, an automatic approval review, scheduled tasks and voice input. Community plugins live under the dsh-plugin topic on GitHub. Treat any plugin you did not write as code that can run on your computer, and read what it does before installing it.

There are also modes, which change what the assistant starts with. The vendor's documentation names Standard and Creator presets. Creator mode lets you build a plugin by describing it in chat. If you switch on Coding Tools in the general settings, two more presets appear, named PTC and Minimal; we have not confirmed what PTC stands for on the vendor's pages, so we will not guess. Skills are saved instructions the assistant can load when a task calls for them.

Step 10: scheduled tasks

Harness can run recurring work on a schedule. DeepSeek's own example is a weekly report: "Every Friday at 17:00 Beijing time, summarize this week's project notes in a weekly report." The page marks scheduled tasks as experimental. The pages we read do not say whether a task still runs when Harness is closed, so assume your computer must be on and Harness running, and test with a harmless task before you rely on one.

What it costs

The app is free. It is MIT-licensed, which means you can use it for any purpose without paying DeepSeek for the software. What you pay for is the model.

On DeepSeek's own service, prices are quoted per million tokens, where a token is, in DeepSeek's words, "the smallest unit of text that the model recognizes", such as a word, a number or a punctuation mark. DeepSeek's price list, checked 11 November 2026, shows two models, and prices depend on the time of day:

DeepSeek's own price list, checked 11 November 2026. Prices are per million tokens, for input text that was not already cached. Cached input is cheaper, and the vendor reserves the right to change prices.

The vendor states that off-peak rates are half of peak rates, and that peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Every other hour, including weekends, is off-peak. Because you are billed for the text the assistant reads and writes, a long session in a big project costs more than a short one. Set a small top-up on your DeepSeek account so a runaway task cannot spend more than you intended.

Privacy: where your data goes

Harness itself runs on your computer. But a model is a service. If you use a DeepSeek key, the instructions you type and the file contents the assistant reads are sent to DeepSeek's servers to be processed. The same is true of any hosted provider you pick instead. If that is not acceptable for your files, you have two choices: use a provider you trust more, or point Harness at a model that runs on a machine you control through the custom model option. Do not put customer records, passwords or anything regulated into a project folder you give to a hosted model.

The separate "Super Agent" prototype

You may see DeepSeek's Super Agent mentioned. It is a separate repository called dsh-superagent-prototype that runs on top of Harness. Its README is blunt: "Research prototype for reference only, not an official DeepSeek product," to be treated "as research code, not production-ready software", and some designs in the accompanying article are "directions for future development" that are not built. It is a proactive agent for one GitHub repository that opens a work item when an allowed person asks for something and answers with a pull request. Setup has many steps. For a small business, treat it as reading material, not a tool to deploy.

When a small business should use it, and when not

Reasonable uses: a developer you employ or contract wants a free, flexible coding assistant and a low model bill; you want to experiment with an AI helper on a copy of a site or script, with backups; you want to see how coding agents work before paying for one.

Skip it for now if:

• nobody involved is comfortable with a terminal or version control;
• the work touches customer data, payment systems or anything you cannot restore;
• you need guaranteed stability or someone to call. This is a developer preview with no support promise;
• your real bottleneck is not code. Most small businesses we advise are held back by follow-ups, bookings or paperwork, which a coding tool does not touch. Start with our plan finder instead.

If you want to see someone click through the app first, there is a video walkthrough by James NoCode that we found useful: watch it here. We did not use it as a source for any fact in this guide.

Compare it with the alternatives

We keep a sourced comparison of coding agents on the harness comparison page, and a table of every plan and price in AI coding agent plans compared. If you prefer a paid product with a company behind it, read our guides to Devin and OpenCode. If you want a very small tool you can reshape, see Pi. Model prices are on the AI model tracker.