DeepSeek Harness desktop agent enters public preview as open source on macOS and Windows

DeepSeek has moved its agent tooling, DeepSeek Harness, into open public preview. According to the announcement on DeepSeek's site, which reached the front page of Hacker News on October 2, the software runs as a desktop application on macOS and Windows, and the project's code is published openly. The same harness can also be launched as a web UI directly from code.

A general-purpose desktop agent

DeepSeek presents Harness as a do-everything assistant for local work rather than a single-purpose tool. The announcement groups its capabilities into a few buckets: everyday tasks such as organizing files, analyzing data and drafting documents or slide decks; coding chores spanning repository exploration, bug fixes, feature work and running tests; research jobs that involve finding information, verifying facts and citing sources; and background operations that execute scripts, batch-process files and report progress over time.

Interaction mixes natural language with familiar agent-interface conventions. Users describe what they want, invoke slash commands, and reference files or sessions with @ mentions.

The plugin-first architecture

The central technical claim is architectural. Per the announcement, Harness builds on Cordis's "everything is a plugin" design, meaning tools, skills and even parts of the interface are meant to be extended through plugins rather than shipped as fixed features. Existing plugins can be installed, and new ones can be created without leaving the app via a chat-driven "Creator mode."

One extension DeepSeek highlights is a Scheduled tasks plugin, which can kick off tasks at arbitrary times and run recurring work on a timetable, with progress visible and tool call details available for inspection. Output such as documents, spreadsheets and code can be previewed inside the app and then refined through follow-up conversation.

What developers get

Debugging support is built in: the announcement points to execution traces and detailed runtime information for troubleshooting tool calls and task execution. The on-ramp is deliberately light. With Node.js installed, the web UI starts with a single command:

npx @deepseek-ai/dsh web

Alternatively, the full source can be cloned from the repository under DeepSeek's GitHub organization and built following its setup instructions. The packaged desktop build targets macOS and Windows, according to the release's title.

Why it matters

Agent runtimes are shaping up to be the next contested layer of the AI stack, and DeepSeek, a lab best known for releasing capable open-weight models, is now competing at that layer with an open-source product rather than a closed, hosted one. That choice matters for software whose job is touching your filesystem: an auditable codebase gives developers and security teams a way to verify what an agent that organizes files and executes scripts actually does.

The plugin model is also an ecosystem bet. If harnesses become the place users spend their time, the platform with the richer extension community gains the same compounding advantage editors and browsers have historically enjoyed, and letting users write plugins through plain conversation lowers the barrier for that community to form.

Questions remain that the announcement does not answer, including which model powers the agent, what license governs the code, and how the desktop build gates sensitive actions like script execution. As with any preview, features may also change before a stable release. Even so, an open-source desktop agent from one of the most closely watched AI labs is a notable entry in a category where major labs have largely shipped closed products, and it gives developers an early look at how DeepSeek thinks agent tooling should be built.