DeepSeek Harness: open harness enters public preview

DeepSeek has released DeepSeek Harness into public preview. The harness is available worldwide and is distributed as open-source software. It works with files, data, and code.

It is built on the Cordis architecture, which the developers describe with the formula "everything is a plugin". The basic set of capabilities is extended by pluggable modules, including ones written for a specific task, and it is plugins that extend what agents can do. The official page lists among them agent teams, auto-approved review, voice input, a terminal, and scheduled task runs, with some marked as experimental.

How to launch

There are two ways to launch the harness. There is a desktop application for macOS on Apple silicon and for 64-bit Windows, or you can run a web interface straight from the code. Node.js is all you need to install it: the web version starts with the command npx @deepseek-ai/dsh web, and the source code can be cloned from the repository.

The agent works with files and documents, analyzes data and spreadsheets, and prepares presentations. In programming, it explores repositories, finds and fixes bugs, adds features, and runs tests. Searching for information with fact-checking and links to sources is listed as a separate area. Long operations such as running scripts and batch processing files run in the background with progress tracking.

Plugins and Creator mode

You can extend the harness in two ways: install ready-made plugins or create your own right in the chat, in Creator mode. Through it, tools, skills, and interface elements are added. For developers, debugging tools are provided, where execution tracing and details about the runtime environment are visible.

The project is distributed under the MIT license. The official page calls it open, reusable, and composable infrastructure, and invites users and plugin authors to grow the harness ecosystem.

Translation editor: Салтанов А.