DeepSeek Harness Tutorial: How to Build an AI Coding Agent

DeepSeek Harness is an open-source AI coding agent that can write code, edit files and run terminal commands. Unlike closed-source coding agents such as Claude Code and Codex, its source code is publicly available.

This DeepSeek Harness tutorial explains the difference between an AI model and an agent harness, its plugin-based architecture, installation process, runtime modes and practical applications.

Watch the complete DeepSeek Harness installation and website-building demonstration. 

What Is DeepSeek Harness?

DeepSeek Harness, also known as dsh, is an open-source runtime that connects AI models to project files, development tools and terminal commands. It is not a new AI model. Instead, it provides the working environment an AI model needs to perform practical coding tasks.

The platform is currently in developer preview, and compatibility-breaking changes may occur as it develops. For now, it is best suited to learning, experimentation and controlled testing rather than critical production work.

AI Model vs Agent Harness

An AI agent can be explained using a simple formula:

Agent = Model + Harness

The model is responsible for understanding prompts, reasoning about a task and deciding what to do next. The harness converts those decisions into real actions.

A harness manages questions such as:

• Where are the project files?
• Which tools are available?
• Can terminal commands be executed?
• Should the agent request permission?
• How should sessions be saved?

A Simple Real-Life Example

Think of the AI model as a driver. The driver may know how to drive, but cannot travel anywhere without a car.

The car provides the steering wheel, brakes, engine and navigation system. In this example, the driver represents the model, while the car represents the harness.

Coding agents such as Claude Code and Codex also provide a runtime between the model and the project. DeepSeek Harness performs a similar function, but its source code is available under the MIT licence.

Understanding how models, tools and workflows operate together is an important part of building modern AI agents. Learners can explore these concepts practically through the AI Agent Course at GALTech School of Technology.

Why Is DeepSeek Harness Different?

The main design principle of DeepSeek Harness is:

Everything is a plugin.

Its models, tools, skills, sessions, sandboxes, storage systems, loops, scheduling features and user interface can all be composed as plugins.

DeepSeek Harness uses an open-source plugin kernel called Cordis. It manages plugin mounting, dependencies and communication between different services.

This modular architecture offers greater flexibility. A model or tool does not have to remain a permanent part of the coding agent. Users can connect DeepSeek, Anthropic, OpenAI, Kimi, MiniMax or another supported provider. Organisations with their own AI gateway can also configure a custom provider.

How to Install DeepSeek Harness

Before beginning, make sure Node.js is installed on your computer.

The quickest way to launch DeepSeek Harness is to open a terminal and run:

npx @deepseek-ai/dsh web

This command starts the local Web UI. By default, it becomes available at:

http://127.0.0.1:3080

If the interface does not open automatically, copy this address and paste it into your browser.

Developers can also clone and build the source code:

git clone https://github.com/deepseek-ai/deepseek-harness.git

cd deepseek-harness

pnpm install

pnpm run build

pnpm dsh web

The source installation is useful for developers who want to inspect the project architecture or understand how its components work.

Setting Up the Web UI and API Key

When DeepSeek Harness opens for the first time, it displays an internal testing notice. This confirms that the platform is still in its testing phase.

The setup process asks you to enter a DeepSeek API key. After adding the key, select Save and Continue to access the Web UI. You can also skip this step and configure the key later.

To edit the API key or choose another model, open:

Settings → Models

The interface allows users to switch between available DeepSeek models, add another supported provider or configure a compatible custom model provider.

Models, Plugins and Interface Settings

DeepSeek Harness provides several customisation options. Users can:

• Select an agent preset
• Change permission settings
• Choose English or Chinese
• Select a light, dark or system theme
• Edit model API keys
• Switch between model providers
• View available plugins

Different versions of a selected model can be chosen from the interface. Users can also adjust the model’s effort level based on their requirements.

DeepSeek Harness Permission Options

DeepSeek Harness provides different permission levels for controlling what the coding agent can do:

• Read Only: The agent can inspect files without modifying them.
• Workspace Write: The agent can work with files inside the selected workspace.
• Full Access: The agent receives broader access to the system.

Greater access gives the agent more control. Therefore, users should select permissions carefully according to the task.

Understanding the Agent Presets

DeepSeek Harness offers four main agent presets.

Standard Mode

Standard mode provides a complete coding-agent setup. It is suitable for common coding tasks involving files, tools and commands.

PTC or Code Mode

Code mode includes Standard mode capabilities and allows the model to coordinate multiple tool calls using model-generated TypeScript code.

Generated code may receive authority similar to Bash commands, so this mode should be used carefully in a trusted testing environment.

Minimal Mode

Minimal mode is another available preset that users can select according to their task requirements.

Creator Mode

Creator mode provides a different runtime configuration for users who want to experiment with the agent environment.

Users can switch between these presets based on the type of work they want the agent to perform.

Choosing a Project Workspace

To work on a project, select Choose Workspace and choose the required folder.

Once connected, the agent can access the project files according to the selected permission level. Users can search available workspaces, change the view and add new workspace folders.

This makes it possible to use DeepSeek Harness with an existing coding project instead of always creating a new project.

Building a Fitness Website with DeepSeek Harness

The video demonstrates the tool using a simple instruction:

Create a website for a fitness club.

After receiving the prompt, the agent begins reasoning, selects the required tools and creates the website files.

The result is a professional-looking, single-page fitness website containing:

• A homepage section
• An About section
• Fitness classes
• Trainer information
• A consistent colour scheme
• A polished visual appearance

The complete page is generated in approximately one minute. This example shows the difference between receiving a code snippet and allowing an AI coding agent to work directly inside a project environment.

Understanding the Trajectory Feature

During the website demonstration, DeepSeek Harness records the complete process using its Trajectory feature.

Trajectory allows users to inspect:

• The system prompt given to the model
• Recorded reasoning events
• Tools called by the agent
• Subagent scheduling
• Completed tasks

Instead of seeing only the final website, users can examine the actions and tool calls involved in producing it.

Other Practical Applications

DeepSeek Harness is not limited to building websites. Users can connect an existing application or coding workspace and ask the agent to:

• Read and analyse project files
• Evaluate existing code
• Identify possible errors
• Suggest corrections
• Modify project files
• Work on larger coding applications

These workflows demonstrate how agentic AI can support software development and automation. Those interested in creating more advanced intelligent applications can explore GALTech’s AI Engineering Course.

Conclusion

DeepSeek Harness demonstrates how an AI model can become a functional coding agent when connected to the right runtime.

Its open-source architecture, model flexibility, permission controls, agent presets, workspace access and Trajectory feature provide greater visibility into how an agent completes a task.

Begin with a test project, select the appropriate permissions and review the actions performed by the agent. You can also explore the complete range of AI and technology courses at GALTech School of Technology.