DeepSeek Harness

DeepSeek Harness (command-line name dsh) is an open-source agent harness developed by DeepSeek AI for running and organizing AI agents. The project uses an everything-is-a-plugin architecture and is built on Cordis. The design ideas behind Cordis are described in the paper "A Programming Paradigm for Spatiotemporal Composability".

⚠️ Developer preview: The project is evolving quickly, and updates may include breaking changes. Before running it, read the safety notes in the repository (SAFETY.md).

Key Features

• Plugin-based architecture: Core capabilities are organized as plugins, which makes them easy to extend and combine.
• Web UI: Start it with dsh web. The default address is http://127.0.0.1:3080. When started locally, it opens your browser automatically; when started over SSH, it only prints the address. Add the --no-open flag to run just the server without opening a browser.
• Plugin ecosystem: Once a plugin repository has the dsh-plugin topic, it can be discovered on GitHub.
• Desktop support: The repository's development docs and Make targets mention both Web and Desktop application forms.
• Multilingual docs: The README and contribution guide are available in English and Chinese.

Target Users and Use Cases

• Developers who want to build or study a runtime framework for AI agents.
• Community contributors who want to customize agent capabilities through the plugin mechanism, and develop and share dsh plugins.
• Technical users who want to try DeepSeek's official agent toolchain.

Installation and Running

Run via npm (Node.js must be installed first):

npx @deepseek-ai/dsh web

Run from source:

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

During development, you can use pnpm run dev:web. It builds, starts the server in a single terminal, and rebuilds the client after source changes.

Technical Requirements and Notes

• Requires a Node.js environment. Running from source also requires pnpm.
• The repository is mainly written in TypeScript, and also contains Python and native directories.
• Licensed under MIT. Third-party dependencies and their licenses are listed in THIRD_PARTY_NOTICES.md.
• It is in developer preview, so APIs and behavior may change at any time. It is not recommended for production environments with high stability requirements.
• Documentation site: https://deepseek-harness.github.io/deepseek-harness/; for community discussion, use GitHub Discussions and Discord.