Get Started with DeepSeek Harness in 5 Minutes: Launch the Web UI and Run Your First Agent

Introduction

The article shows how to quickly get DeepSeek Harness up and running, covering installation, Web UI launch, model configuration, workspace selection, executing a first task, and the sandbox permission system.

1. Start the Web UI

1.1 One‑click launch

After installing Node.js, run the following command:

# Need Node.js installed
npx @deepseek-ai/dsh web

The command prints an address, typically http://127.0.0.1:3080. Opening this URL in a browser displays the Web UI.

1.2 Source installation

Clone the repository and install dependencies:

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

These steps also start the Web UI.

2. Configure the model

When the UI loads for the first time, it prompts you to enter your DeepSeek API key, which you obtain from the DeepSeek platform. The key is write‑only; after saving, only a masked descriptor is shown, and the key is stored locally at $DSH_HOME/.credentials.yaml (or ~/.dsh/.credentials.yaml on macOS).

3. Choose a workspace

The workspace defines the directory that the Agent can read and write. Click Choose Workspace and select the project folder where you ran dsh. If no workspace is selected, the input box is greyed out.

The UI offers four modes; the default Standard mode works for most cases.

4. Run the first task

Create a new session and enter a query, for example:

What is the difference between Workspace Write, Full access, and Read Only?

After pressing Enter, the Agent begins working, reading files, executing commands, and advancing the task. The article includes a diagram of the Agent’s execution flow.

5. Permission approval

At the bottom left of the UI, a file sandbox shows three permission levels:

Read‑only : view only.

Workspace‑write : modify files inside the selected workspace.

Full access : read and write anywhere on the system.

When an operation exceeds the current permission, the sandbox blocks it and returns a message such as [sandbox: file access denied under <mode> mode]. You can request a temporary elevation for a specific command; the UI asks for approval, and the request is one‑time only. If you reject, the Agent stops and explains why.

The approval mechanism itself is pluggable via dsh-user-approval, and the policy can be set to ask (prompt) or never (auto‑reject). A “fail‑closed” setting ensures that without a response, the operation is denied.

6. Summary

Start: npx @deepseek-ai/dsh web → http://127.0.0.1:3080 Configure model: Settings → Model → Enter API key → Save (no restart needed)

Select workspace: must be chosen or the input box stays disabled

Run task: type a query, the Agent reads files, runs commands, and progresses the plan

Permission approval: sensitive actions trigger a prompt, keeping you in control