DeepSeek released the preliminary version of its open-source agent application, DeepSeek Harness, version v0.2, on September 29. For the first time, users can obtain ready-to-run desktop installers for macOS on Apple silicon processors and 64-bit Windows via the official Harness website.
Earlier this week, the publication Pandaily reported the appearance of preliminary Electron builds distributed from the DeepSeek update server, even before the official announcement. Version 0.2 represents the official desktop release and includes several features beyond just a wrapper.
Harness initially appeared as an open-source developer preview, v0.1, on August 13. The most significant addition is the plugin management section. Since Harness is built on the principle that everything is a plugin, installing a plugin no longer requires using the command line: users enter the npm plugin package name into a dedicated field and can disable or delete plugins while viewing their description and source in one place.
In experimental creation mode, users can describe the desired plugin in a conversational format, after which Harness automatically generates or modifies it; the new plugin is immediately activated, saved, and displayed in the manager. DeepSeek notes that about 60% of Harness users, including those using the official API model, already use third-party plugins. The company plans to create an official plugin marketplace and ensure a more stable plugin API with fewer compatibility-breaking changes.
Version v0.0 also includes a built-in automation task plugin, which users activate through the plugins page. This tool allows routine operations to be turned into scheduled tasks, providing execution history, frequency setting options, and instruction editing. In a practical review, the technology resource ifanr found that scheduled tasks continue to run even after the application is closed. Furthermore, the plugins page lists four experimental features: agent commands, automatic authorization checks, automation tasks, and voice input, which loads a local speech recognition package of about 1 GB.
The update has also changed how files are displayed, content previews, and code changes. Users can provide Harness with documents, spreadsheets, or PDF files for material organization, data analysis, chart generation, or presentation creation, or ask it to make changes to a project and execute code. Generated files and each step of code changes are recorded in the dialogue; the sidebar shows files and differences (diffs), and results can be opened in local applications. Users also have the option to adjust the level of agent workflow disclosure.
DeepSeek asserts that Harness has become the most popular coding agent among users of its official API, according to daily active user and session metrics, while emphasizing that the product is still in its early stages of development. The company's plans include strengthening the sandbox and security, improving agent commands, personalized long-term memory, browser and GUI automation, remote and mobile device usage, as well as session sharing and multi-user collaboration. The company also intends to optimize future models to improve performance within Harness.