DeepSeek packages its unaudited Harness agent runtime as a Windows and macOS desktop app

The key word in DeepSeek's terms of use is "available". According to the terms, the services can run model-generated code and commands, load third-party plugins, and access files, credentials, processes and networks made available to them ^[12]. Harness does not set its own limit in that sentence. The user sets it, by deciding what the account running the app can reach.

That decision also shapes what gets logged. A chat model's input is what the user typed. An agent's input includes the documents it opened and the commands it ran. I'd assume any file Harness reads on DeepSeek's own models can be logged, since DeepSeek's list for official-model sessions includes authorized content and tool calls ^[16].

Before the desktop build, DeepSeek's documented web interface required installing Node.js and running a command, so every user had at least opened a terminal ^[7]. Installing Node.js was never a security control, but it did a fair impression of one. The desktop downloads cover Windows and Apple-silicon Macs on macOS 13 or later ^[3].

The design pulls in two directions. DeepSeek pitches Harness for work that goes on without the user: running scripts, analyzing spreadsheets, processing files in the background ^[5]. The safety notice asks users to grant only the permissions a task needs and to review plugins and proposed commands ^[14]. Following that advice literally means watching the agent the whole time. On a laptop that holds client files and cloud credentials, I would run it under a separate OS account with one project folder exposed, and keep background tasks off until the preview label comes off.

Some of the engineering deserves credit. Cordis, the plugin layer, lets users install or write extensions for tools, skills and interface features ^[6]. The runtime is MIT-licensed and still a developer preview, according to its GitHub repository ^[17]. Because the code is open, anyone can read what a tool call actually does before deciding to trust it. The safety documentation states the failure modes in plain words, and that is to DeepSeek's credit ^[13].

Runtimewire frames the release as DeepSeek extending its open-source work from models into the agent software that uses them ^[15]. The record so far fits that, within limits the publication itself sets out. The release does not show how many people will use the app, whether they will keep it pointed at DeepSeek's models, or how much work they will trust it to do unattended ^[11].