DeepSeek Releases Desktop Harness Agent for macOS and Windows

The official Harness website now provides downloadable packages for macOS and Windows. Users who install the desktop version and log in receive a 6 yuan credit, valid until Oct 6. The desktop client can stay resident on a local computer, allowing it to run in the background, execute scheduled tasks or install new capabilities like Codex, ifanr reported.

Login uses a DeepSeek open platform account and its balance, and does not sync chat conversations from the DeepSeek app, according to the report. After signing in with an account or API key, users see an interface similar to the terminal version, including four Agent presets: Standard mode, PTC mode, Minimal mode and Creation mode, along with plugins and custom third-party models.

DeepSeek's description of Harness has shifted. In the official introduction for the desktop version, coding is presented as only one part of the product; daily work is described as the more suitable scenario. According to ifanr, users can give Harness a Word file to organize materials, an Excel file to analyze data and generate charts, or a PDF to extract information, and then produce a presentation. Harness can browse Word, Excel, PDF, Markdown and code files in a Workspace and leave processed results in the same folder. Office and PDF previews are built into Harness; the 0.2.0-rc.1 release on Sept 28 optimized text selection and preview for Office and PDF files.

Ifanr tested the desktop version with a folder containing PDFs, Word documents, Excel files, Markdown, images and miscellaneous materials. The instruction was to read all materials, organize important information, verify data and generate a report outline and a presentation. The test showed Harness running at nearly 300 tokens per second and using two sub-agents, the report said. It decided which files to read and which tools to call. Generated files and code changes from each round are displayed in the conversation; users can preview file contents in a sidebar, view code differences, continue asking for modifications or open files in local applications for further editing.

The desktop version also adds Automation as a plugin that users can enable as needed. According to ifanr, a user can tell Harness, 'Every Friday at 5 p.m., read the new materials added to this project folder this week and compile a weekly report.' Harness creates a recurring task, and a separate interface lets users view and modify when the task runs, when it will run next and whether it succeeded previously. Even if the desktop client is closed, the scheduled task is not canceled. Automation is currently a lab feature and must be enabled by clicking into the plugin page. The plugin page lists four features—Agent Teams, automatic authorization review, automation tasks and voice input—all requiring manual activation. Voice input requires downloading a local speech recognition package of about 1 GB, including the model, runtime and download cache. The model, named SenseVoiceSmall, is a lightweight multitask speech understanding model released by Alibaba's Tongyi Lab, DAMO and the FunAudioLLM and FunASR teams. Ifanr said its voice recognition tests were fast and accurate.

The desktop version and the new release add a full plugin management page. Plugins that previously required command-line installation can now be installed by entering an npm package name, and installed plugins can be viewed by source, disabled or uninstalled. DeepSeek said in its release that about 60 percent of Harness users, based on official model API users, have used third-party plugins, according to ifanr. DeepSeek plans to build an official plugin marketplace. In Creation mode, users can also ask Harness to write a DSH plugin. In one example, the instruction 'write a manuscript checker plugin' led Harness to read the plugin development Skill, check available interfaces, write JavaScript and call the plugin manager to install it; about 10 minutes later, the manuscript checker appeared in Harness's interface.

DeepSeek also outlined future work in its official article. The company plans to make the plugin marketplace more active and secure, strengthen sandbox isolation, improve multi-agent collaboration through Agent Teams and add personalized long-term memory. Future desktop versions will also add automation for browsers and other desktop GUI software—Computer/Browser Use—and support remote access and mobile deployment.