DeepSeek brings its open-source agent runtime to Windows and Mac

Why it matters

DeepSeek is extending its open-source work from models into the agent software that uses them. Desktop access broadens the potential audience, while local file permissions and model-provider data handling become central adoption questions.

DeepSeek has made its open-source Harness agent runtime available as a desktop app for Windows and macOS, giving users a graphical route into software previously launched through a command line and browser. DeepSeek's Harness account on X posted the availability notice on September 30th; the company's download page lists Windows and Apple-silicon Macs running macOS 13 or later.

The product comes from the AI company founded by Liang Wenfeng, whose own technical path began in machine vision. Zhejiang University says Liang completed his master's thesis at its College of Information Science and Electronic Engineering, after a supervisor introduced him to machine vision. Liang later co-founded High-Flyer, the quantitative investment firm that helped build DeepSeek's research operation. The desktop rollout extends that research organization into software people can install and use on their own machines; DeepSeek has not tied the release to a named product lead.

Harness is an agent runtime: a layer that connects a language model to tools for acting on files, code, and other systems. DeepSeek says it can help organize documents, analyze spreadsheets, write and test code, conduct research, run scripts, and process files in the background. Its design is built around Cordis, a plugin architecture that lets users install or create extensions for tools, skills, and interface features. The project is released under the MIT license and remains in developer preview, according to the official GitHub repository.

The desktop version lowers the setup hurdle. Before the app, DeepSeek's documented web interface required users to install Node.js and run a command; the desktop downloads bundle an application rather than asking users to open a terminal and browser themselves. That packaging brings the same runtime closer to people who want an agent to work across local documents and projects, rather than only respond in a chat window. DeepSeek says the app can edit local files and keep tasks running in the background.

The desktop label does not mean all model processing stays on the computer. DeepSeek's terms of use describe two modes: official DeepSeek models, which require an account, and custom models configured with an API the user obtains. For custom models, DeepSeek says the model provider supplies the model service. Its privacy policy says inputs sent through custom-model calls go directly to that provider, which applies its own data policies. For official-model sessions, DeepSeek says it collects session logs that can include user inputs, authorized or uploaded content, model outputs, tool calls, plugin configuration, and conversation data.

That distinction matters for a desktop agent with access to local work. Harness is designed to take actions, not simply draft answers: DeepSeek's terms say the services can run model-generated code and commands, load third-party plugins, and access files, credentials, processes, and networks made available to them. The project's safety notice says the developer-preview software has not undergone a security audit and warns that mistakes, malicious input, or untrusted plugins can damage files or disclose data. The notice recommends using only the permissions needed and reviewing plugins and proposed commands.

Those warnings put the release in the same practical category as other coding and computer-use agents: useful precisely because they can operate beyond a chat box, and deserving of deliberate access controls for the same reason. DeepSeek's pitch is an extensible, open-source environment that can be adapted through plugins and configured with official or user-supplied model services. The release adds a desktop front end; it does not establish how many people will use it, whether they will keep it pointed at DeepSeek's own models, or how much work they will trust it to perform unattended.

For Liang, whose company grew out of an organization built around quantitative research and computing, Harness is another outward-facing layer for DeepSeek's technical work. Its desktop distribution puts the agent runtime in front of a broader audience than developers willing to install packages and launch a local web interface. The immediate test is operational: whether a preview agent that can edit files and execute commands proves dependable enough for everyday work, and whether its plugin model gives users a reason to adopt DeepSeek's runtime rather than an alternative.