- Guides
- Tool guides
- How to use DeepSeek Harness
How to use DeepSeek Harness
DeepSeek's open-source agent works on documents, spreadsheets and code from a desktop app or your browser. It is a developer preview with no security audit, so here is how to install it and give it a safe place to work.
We have no affiliate deal with DeepSeek, and nobody paid to appear on this page.
What DeepSeek Harness is
DeepSeek Harness is a free, open-source AI agent that works on files on your own computer. DeepSeek calls it "an open-source agent harness", released under the MIT license, and its command is dsh 1. You give it a folder and a task in plain language; it reads and edits files, runs commands, splits work across sub-agents and keeps a plan 4.
It is not limited to code. DeepSeek's product page shows it producing Word, Excel, PDF, HTML and Markdown files, with a preview of each result, alongside writing code and changing repositories 3. Every capability is a plugin, and it can write new plugins for you in a Creator mode. It also runs scheduled tasks and shows a trace of each run 3.
The software costs nothing. The model behind it does: you add a DeepSeek API key and pay per token, usually cents per task (see what it costs). Treat it as a preview. The README says it is in developer preview and "THERE WILL BE COMPATIBILITY-BREAKING CHANGES" 1, and the safety notice says it "has not undergone a security audit" 2.
It fits founders and operators who want a desktop agent for reports, spreadsheets and small automations without a monthly subscription, and who are comfortable checking its work. If you want a polished product with support, a paid agent such as Claude Code is the safer pick; our Claude Code install guide covers it.
Install it
Most people should download the desktop app. If you have Node.js, one command starts it in your browser instead. Either way, you then add an API key before the agent can do anything.
Option 1: the desktop app
- Download the right buildDeepSeek's Harness page offers two: macOS for Apple chips (a .dmg file) and Windows 64-bit (a .exe file) 3. No build for Intel Macs is listed.
- Install and open itInstall it like any other app, then open DeepSeek Harness.
Option 2: the web UI with npx
Install Node.js first; the README does not name a minimum version 1. Then open a terminal in the folder you want the agent to work in and run:
npx @deepseek-ai/dsh web
This starts the web UI at http://127.0.0.1:3080 and opens it in your default browser. Add --no-open if you don't want a browser tab to open 1. The server runs on your machine; close the terminal to stop it.
Option 3: from source (developers)
The README lists these steps for building from the repository 1:
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web
Add your API key
- Get a keyCreate one on DeepSeek's platform under API keys 11. Usage is charged against a balance you top up there 10.
- Paste it into HarnessOpen Settings, Models, enter the key on the DeepSeek card and save. It works right away, with no restart 4.
The key is write-only in the app and stored in a local credentials file 5. DeepSeek is not the only option: Add model provider lists built-in third-party providers including Anthropic, OpenAI, Moonshot (Kimi) and Z.ai (GLM), and a Custom model API option connects a company gateway or a self-hosted model 5.
Set up a safe workspace
Give the agent one fresh folder that holds only copies of the files it needs. Harness works inside a workspace you choose, and it can read and edit files there and run commands 4. DeepSeek's safety notice is blunt that its sandboxing and approval prompts "do not guarantee isolation or prevent damage" 2. A folder of copies limits what a bad run can break.
- Make a sandbox folderFor example
Harness-sandboxin your Documents folder. Copy in, don't move, the spreadsheet or documents for your first task. - Start from that folderIf you use the web UI, run the
npxcommand from inside it. Thedshprocess uses the folder it was started in as its default location 4. - Choose the workspaceClick Choose workspace, add the sandbox folder and select it. The chat box stays locked until a workspace is selected 4.
- Read every approvalThe app asks before operations that need approval under its permission policy 4. Read the command or file change before you allow it.
For anything sensitive: DeepSeek recommends running with the least privileges possible, preferring "a disposable virtual machine, container, or dedicated environment", keeping backups, and not exposing credentials or data you can't afford to leak 2. Keep client files and password-filled folders out of the workspace.
Your first tasks
Start with a real but low-stakes file task, then check the trace to see exactly what the agent did. DeepSeek's own quickstart opens with a read-only request on a code project 4. For non-developers, a spreadsheet is the better test, which is how we framed it when we covered Harness in our newsletter on 4 November.
A spreadsheet summary
Put a sales or expenses export in the sandbox and ask for an analysis plus a document you can forward:
Open q3-sales.xlsx. Find the three products whose revenue changed most from one month to the next, explain each change in two sentences, and save a one-page summary as q3-summary.docx.
Harness previews the files it creates 3, so you can read the summary in the app and ask for changes in the same chat ("shorter", "add a table of the top ten"). Check every number against the sheet before it goes anywhere. For more spreadsheet prompts, see our guide to AI for Excel.
A folder of documents
Read every PDF in the contracts folder. Build contracts.xlsx with one row per contract: client, start date, renewal date and notice period. Flag any contract that renews in the next 60 days.
Turn the notes in meeting-notes.md into a clean Markdown page with decisions, owners and deadlines.
Check the trace
After each task, open the run trace. DeepSeek says it shows the execution trace and detailed run information 3: which tools ran and which files were read or written. It is the fastest way to find why an output is wrong, and to confirm the agent stayed inside your folder.
Plugins and scheduled tasks
Plugins add skills to Harness, and you can ask it to build one for you. Scheduled tasks run a prompt on a timer.
Build a plugin in Creator mode
DeepSeek's architecture is "everything is a plugin" 1. In Creator mode you describe a plugin in chat and Harness builds and installs it; DeepSeek's demo builds a Pomodoro timer this way 3. A timer is a demo. A more useful first plugin turns your meeting transcripts into a list of action items with owners, in the same format every time:
Create a plugin called action-items. When I give it a meeting transcript, it returns a table of action items with owner, deadline and the sentence that created each one.
Developers can write plugins by hand: a plugin is a TypeScript module, loaded from a source checkout 8, and community plugins are tagged dsh-plugin on GitHub 1.
Install plugins from an index
Browsing plugin catalogs is an experimental setting. Open Plugins, find Custom plugin indexes under Experimental plugins and switch it on; a Discover tab appears next to Installed 7. Each install asks for confirmation. DeepSeek warns that "indexes are third-party content", so add only sources you trust 7. A plugin runs with the agent's access to your files, so read what it does before you install it.
Scheduled tasks
You can create a scheduled task in chat or by hand. In chat, just say it:
Every Friday at 5pm, read this week's files in the notes folder and write a one-page weekly report as weekly-report.md.
To set one up by hand, open Automation tasks in the sidebar, choose Create manually, and fill in the name, the instruction, the linked workspace and the run time. You can also pick whether each run reuses one conversation or starts a new one 6. Run now tests a task immediately, and Records lists past runs 6.
Tip: scheduled tasks only run while Harness is running. Closing it stops scheduling until it starts again, and DeepSeek notes that a delivery record does not prove the task was completed 6. Read the output of the first few runs.
What it costs to run
Harness is free; you pay DeepSeek per token, and a typical task costs cents. DeepSeek's API page lists two models, deepseek-flash (DeepSeek-V4.1-Flash) and deepseek-v4-pro, both with a 1 million token context 10.
| Price per 1M tokens | deepseek-flash | deepseek-v4-pro |
|---|---|---|
| Input, cache miss (peak / off-peak) | $0.30 / $0.15 10 | $1.32 / $0.66 |
| Input, cache hit (peak / off-peak) | $0.006 / $0.003 | $0.044 / $0.022 |
| Output (peak / off-peak) | $1.20 / $0.60 | $3.96 / $1.98 |
| Reads images | Yes | No |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; all other hours, weekends and Chinese public holidays are off-peak 10. For US users, that means US business hours fall in the cheaper window.
A worked example at peak rates: a task that reads 1 million fresh tokens and writes 100,000 costs about $0.42 on Flash ($0.30 + $0.12) and about $1.72 on Pro. Agents re-read the same context many times, and repeated input bills at the far cheaper cache-hit rate. Fees come out of your topped-up or granted balance, and DeepSeek reserves the right to change prices 10.
Tip: start with Flash. Switch to Pro for a task Flash gets wrong, not by default.
Risks and limits
Use Harness for low-stakes work on copies of your files until it leaves preview. DeepSeek calls it "experimental developer-preview software" that "must not be treated as secure or production-ready" 2.
- It can damage files. Harness can run model-generated code and commands, load third-party plugins, and reach the network, processes, credentials and files. Bad output, bugs or malicious input could modify or delete files or leak data 2.
- No security audit. DeepSeek says not to rely on it "as the sole security control for untrusted workloads" 2. Untrusted content includes web pages and documents from strangers that the agent reads.
- Breaking changes. Settings, plugins and workflows may stop working between versions 1. Don't build a client process on it yet.
- Experimental parts. Agent teams and voice input are marked experimental on DeepSeek's page 3, as are plugin indexes 7.
- Your data goes to the model you connect. Files the agent reads are sent to that provider to be processed. With a DeepSeek key, that is DeepSeek's API; see DeepSeek vs ChatGPT for how DeepSeek compares on privacy and plans.
- No warranty. The notice says to use it "at your own risk" 2.
If you want to try it on real work, the most reassuring route is the one DeepSeek suggests: a separate virtual machine or a spare computer, with backups 2. For how agents gain reusable skills more broadly, see our explainer on agent skills.
Questions people ask
Is DeepSeek Harness free?
The software is free and open source under the MIT license. You pay for the model through a DeepSeek API key, from $0.30 per million input tokens and $1.20 per million output tokens on Flash at peak, half that off-peak.
Does DeepSeek Harness run on an Intel Mac or Linux?
The desktop app comes for Macs with Apple chips and for 64-bit Windows; no Intel Mac build is listed. On other machines, you can run the web UI with npx @deepseek-ai/dsh web if Node.js is installed.
Can I use DeepSeek Harness with Claude or GPT models?
Yes. Settings, Models lets you add built-in providers including Anthropic and OpenAI, or a custom API such as a company gateway. You pay that provider's API rates.
Is DeepSeek Harness safe to use?
It is a developer preview with no security audit, and it can run commands and edit files. Use a folder of file copies, read each approval, keep sensitive data out, and for real work follow DeepSeek's advice to run it in a virtual machine or dedicated environment.
Do I need to know how to code?
No, for the desktop app: you type tasks in plain language and approve its actions. Building plugins by hand and installing from source are developer tasks, but Creator mode lets you describe a plugin in chat instead.
How do I uninstall it?
If the desktop app installed the dsh command, remove it first from the app menu with Manage dsh Command, Remove. Then quit the app and delete it like any other app. DeepSeek's uninstall page lists the data folders to delete if you also want sessions and saved keys gone 9.
Related
New features, explained the week they ship
195,000 readers get Uncover AI's practical AI briefing. We show you what is new and how to use it in your work.
Sources
- DeepSeek Harness on GitHub (README)
- DeepSeek Harness safety notice (SAFETY.md)
- DeepSeek Harness product page
- Harness docs: Use the Web UI
- Harness docs: Configure models
- Harness docs: Session reminders
- Harness docs: Plugin indexes
- Harness docs: Your first plugin
- Harness docs: Uninstall DSH Desktop
- DeepSeek API models and pricing
- DeepSeek API: your first API call