deepseek-ai/deepseek-harness

Overview

DeepSeek Harness (dsh) is an open-source agent harness released by DeepSeek AI. It lets you hand everyday work and development tasks to an agent, such as organizing documents, analyzing spreadsheets, writing code and researching material.

It follows an "everything-is-a-plugin" architecture and runs on top of a framework called Cordis. Tools, skills and even the interface can be attached and detached as plugins, and according to the official page, a "creation mode" lets you build and install new plugins through conversation alone. You can use it as a desktop app or launch a local web UI with a single command.

It is worth a look for developers who want to adapt an agent to the way they work, and for teams that want to study or extend the harness architecture itself. The project has also opened up an ecosystem for sharing plugins under the dsh-plugin topic.

Why it is worth a look

Because even core features such as the agent loop, subagents and the terminal are split into plugins, you can do more than add features: you can reassemble the harness itself. A model developer has released its own harness as open source, and the design of the underlying framework, Cordis, has been published in a paper, which makes it a useful reference for anyone comparing or studying harness designs.

Key features

• You can extend the agent's capabilities by installing tools, skills and interfaces as plugins, or by building your own.
• According to the official page, it offers a "creation mode" that builds and installs plugins through conversation.
• You can schedule automated tasks, such as generating a weekly report at a set time each week.
• In the developer tools, you can trace problems by inspecting the execution trajectory and the history of tool calls.
• It can be run in two ways: as a desktop app or as a local web UI.

What it is used for

• Use it to hand repetitive work to an agent, such as organizing documents, analyzing spreadsheets and preparing reports.
• Use it for development work: understanding project code, fixing bugs and running tests.
• Use it to build tools and interfaces suited to your team as plugins and set up your own agent environment.
• Use it to run long tasks in the background, such as running scripts or processing large numbers of files.

Getting started

1. Install Node.js.
2. Run npx @deepseek-ai/dsh web and the web UI opens at http://127.0.0.1:3080 by default (use --no-open to stop the browser from opening automatically).
3. To run from source, run git clone https://github.com/deepseek-ai/deepseek-harness.git, then run pnpm install, pnpm run build and pnpm dsh web in order from the repository folder.
4. Before running it, read the safety guidance in the repository's SAFETY.md.

What to weigh

• It is in developer preview, and the README states that there will be breaking changes, so the API and configuration may change.
• The project was created in August 2026, so its plugin ecosystem and documentation are still taking shape.
• Because the agent can run shell commands and modify files, check SAFETY.md and review workspace permissions.
• It is MIT-licensed, but the licenses of third-party dependencies need to be checked separately in THIRD_PARTY_NOTICES.md.

Related terms and companies

Harness Harness² AI agent Subagent Skills Coding agent DeepSeek

Sources

• deepseek-ai/deepseek-harness README
• DeepSeek Harness｜共探智能上限

Written with AI assistance from the repository README and GitHub data. Star counts and other numbers are GitHub values as of the date shown; features and licensing can change, so check the repository.