Work with files & code
Read, search, write and edit files precisely — including pattern-based discovery (glob) and content search (regex) across a whole project.
Start here
What is DeepSeek Harness?
DeepSeek Harness (the command is dsh) is an open-source agent harness built by DeepSeek AI. It is not a chatbot that only answers questions — it is a program that does work on your computer: reading and writing files, running shell commands, searching the web, calling third-party tools, and running many agents at once for large jobs.
The core architecture is easy to describe: everything is a plugin. Every capability — including the web interface, language models, file tools and external integrations — is installed as a plugin on top of the Cordis framework. That is why DeepSeek Harness can run as a desktop app, a local web interface, a headless CLI for automation, or a Python SDK you call from your own code.
Its official status is developer preview: it moves fast and behaviour can change between versions. This site documents current behaviour and flags the parts you should re-check after upgrading.
20+
built-in tools for files, terminal, web and documents
3
sandbox modes: read-only, workspace-write, danger-full-access
4+
model providers: DeepSeek, Anthropic, OpenAI, Kimi, GLM, local models
MIT
open-source licence, free to use
Read this before using it on important work
DeepSeek Harness is still experimental and has not been through a comprehensive security audit. Do not hand it full access to production servers or sensitive data on day one. Start with a separate working folder, use read-only while exploring, and review every file change before approving it.
Capabilities
What you can do with DeepSeek Harness
Six groups of work that DeepSeek Harness users do most often — all driven by plain-language instructions.
Run the terminal
Run shell commands, install dependencies, execute tests and read the results. Persistent terminal sessions handle long-running work.
Subagents & parallel workflows
Split large jobs: one subagent researches, one writes, one verifies — or a single workflow script coordinates dozens of agents.
Web research & documents
Search current information on the web, open specific pages, then turn the results into Word reports, PowerPoint decks or Excel spreadsheets.
Extend it yourself
Add skills (SKILL.md files), TypeScript plugins and MCP servers — so your company's internal tools can be called directly by the agent.
Control permissions & safety
Restrict file access with three sandbox modes, require human approval for sensitive actions, and keep a record of every step taken.
How to start
Three first steps to use DeepSeek Harness
You do not need to change anything on your system. If Node.js is installed, DeepSeek Harness runs within minutes.
Make sure the prerequisites are installed
DeepSeek Harness needs Node.js ^22.19.0 or ≥ 24, pnpm 11.28.5 (for source installs), Git, and Python 3.10+ if you use the Python SDK. Check your Node version with
node -v.Run it via npx (fastest route)
A single command starts the local web interface and opens your browser:
Choose a workspace and connect a model
In the web interface, click Choose workspace and point it at your project folder. Then open Settings → Models and paste a DeepSeek API key from platform.deepseek.com. The key is stored in a local credentials file, and you can switch or add other providers at any time.
powershell / bash
node -v
npx @deepseek-ai/dsh web
# server saja tanpa membuka browser
npx @deepseek-ai/dsh web --no-openRequires Node.js 22.19+ or 24+. The web interface runs at http://127.0.0.1:3080 by default.
Beginner tip
Start with a task you can verify in one minute, such as “summarise this folder into a table of contents”. Once you trust the output, scale up. More detail in the usage guide and installation page.
Safety
Understand permissions and sandbox before delegating big tasks
DeepSeek Harness separates two things: how much file access it has (sandbox) and whether risky actions need human approval (approval policy).
| Sandbox mode | File access | Best for |
|---|---|---|
read-only | Read only; writes and edits are refused. | First experiments, reading unfamiliar code, audits. |
workspace-write | Free to write inside the workspace; outside it is refused. | Daily work on a project you chose deliberately. |
danger-full-access | Full filesystem access; no sandbox restriction. | Advanced tasks that genuinely need wide access — with full awareness. |
Alongside the sandbox mode there is an approval policy: ask requests your confirmation for out-of-bounds actions, while never proceeds without asking. A combination of both is called a preset. Two presets ship by default: workspace-write (workspace-write + ask) and danger-full-access (danger-full-access + never). If approval cannot be obtained the request is rejected — the system fails closed rather than quietly allowing it.
A practical rule for teams
Give the agent a dedicated workspace containing only files it may change. Keep credentials outside that folder, enable version control (Git) before large tasks, and make reviewing the diff before approving a habit.
Selected guides
Continue with the guide you need
Installation
Step by step for Windows, macOS and Linux: npx, source builds, profile selection and API key setup.
Usage guide
How to instruct effectively, pick models, set permissions, and use skills, plugins and MCP.
Use cases
Ten real scenarios with copy-ready example prompts you can adapt.
Quick FAQ
The questions we hear most
Is DeepSeek Harness free?
The software is open source under the MIT licence, so it is free to use. What costs money is the model: you use a DeepSeek API key (or another provider) billed per token. Local models via Ollama are an option with no API cost.
Can DeepSeek Harness work offline?
Partly. Tools such as file operations, code search and the terminal run locally. Model calls still need a provider connection — unless you run a local model through Ollama or your own gateway. Web search obviously needs the internet too.
Is DeepSeek Harness suitable for beginners?
Yes, as long as you are comfortable with the idea of a workspace folder and basic terminal concepts. Start with the web interface, read-only mode and small tasks. This site provides an Indonesian-first step-by-step guide to make that easier.
Need help rolling out DeepSeek Harness?
Free consultation with Andi to choose the right setup, prepare a workflow, or review safety before your team adopts it.
Contact Andi — +62 897-6003-363
Need hands-on help?
Free consultation with Andi
If you want to use DeepSeek Harness for work or business but are unsure where to start, contact Andi at +62 897-6003-363 (phone or WhatsApp). The first consultation is free: we discuss your needs, the risks to manage, and the most sensible way to try it.