A free AI employee now works directly on the desktop, with no maze of terminal windows and browser tabs. DeepSeek published the Harness v0.2 preview on September 29 2026 with ready installers for macOS and Windows, turning the old command-line-plus-browser arrangement into a one-click desktop app while carrying past session work across. The architectural framework and official distribution details of this release were obtained from the comprehensive product documentation that DeepSeek publishes on its Harness page.
Installation starts on the official download page, and the desktop build bundles its own runtime so no separate Node install is needed. On Linux the path still runs through the command line; after installing a supported Node version, the npx @deepseek-ai/dsh web command opens a local server with the interface running in the browser. An API key and an account sign-in spend from separate budgets, so confirming which route is selected in settings before starting costly work is essential. These installation steps and the difference between the two billing routes were obtained from the detailed hands-on setup guide that MindStudio published on October 3 2026.
Carrying old sessions across makes a real difference in daily use; previously produced SEO content, an Obsidian vault and landing pages appear ready in the new app on the speaker's screen. Generated files and code changes surface inside the conversation, with previews and diff reviews in a right-hand sidebar and any file openable in its local application. On the document side, Word, Excel, PowerPoint and CSV files can be read, while organizing information, analyzing data, generating charts and preparing slides run in the same session. These document capabilities and the sidebar review flow were obtained from the comprehensive release analysis that Augmenter published on October 2 2026.
Everything is a plugin: manager, teams and voice input
At the center of the design sits the everything is a plugin principle; the model adapter, tool registry, agent loop and even the interface itself run as replaceable parts. The MIT-licensed open-source build also bundles the dsh command and uses a composable plugin system built on the Cordis architecture. The new plugin manager installs from a typed package name, enables one-click disabling and removal, and shows each plugin's description alongside its source. Built in come agent teams, voice input, subagents, web search and shell access.
The community has scaled this architecture fast; the curated plugin list shown in the video has passed 18,000 stars and splits into categories such as memory, interface themes and model providers. The main repository appears past 240,000 stars with 29,000 forks, which makes the project one of the most watched open-source efforts in a short time. Memory-style plugins preserve context across long jobs, and setups keeping a step-by-step todo list while coding can be assembled. These star counts and the category structure of the community list were obtained from the official repository and curated plugin list hosted on GitHub.
Sessions run across four modes: standard mode is now the built-in default for general work, PTC runs programmatic tool calls as Node code in a separate process, and minimal mode offers the lean loop used in DeepSeek evaluations. The experimental creator mode is the most interesting; it builds a new plugin from a written description in chat, saves it automatically and lists it in the manager, while agent presets define the working style up front. The lean loop has a serious side: DeepSeek-V4-Flash-0731 scored 82.7 on the Terminal Bench 2.1 test in that setup. These mode definitions and the benchmark score were obtained from the technical release review that MarkTechPost published on October 3 2026.
Model choice and scheduled tasks
The model side is not locked to one provider; alongside DeepSeek, Anthropic, OpenAI, AWS Bedrock, Google Vertex and Azure plus custom OpenAI-compatible endpoints can be connected. For a free setup, adding a key through OpenCode or OpenRouter runs no-cost models, with the speaker recommending Flash 3.1 class models for that route. Account sign-in users get web search without an extra key, and a searchable model picker simplifies switching providers. This provider variety and the emphasis on in-app plugin creation were obtained from the short release brief that BitInsider published in early October 2026.
The new automation tasks plugin brings a scheduler inside Harness; one-off and recurring jobs can be defined, run history is kept, and repeat frequency plus instructions stay editable afterward. Scheduled jobs surviving restarts can repeat as often as once per minute per the release notes, yet the app must stay open in the background for schedules to fire. Examples such as preparing slides from the week's AI news every Monday morning are set up from a one-sentence description, while step visibility and preview styles ease tracking and debugging.
Security warning, roadmap and quick start
On security there are three sandbox modes: read-only, write-inside-workspace and full access; wider scope asks for approval, and a command does not run when the sandbox cannot be enforced. The speaker gives an explicit warning here; because the harness is a China-based build, private data one is not comfortable sharing should stay out of this environment. It also matters that data flowing through the DeepSeek API is processed under Chinese rules, a point especially relevant for corporate and Europe-based teams. This cautious stance demands putting data classification first while testing a free yet powerful agent.
The roadmap lists stronger sandboxing and security, more effective agent teams, personalized long-term memory, browser and GUI automation, remote and mobile access, and an official plugin marketplace. The quick start finishes in three steps: open a new session, pick the project folder, set the mode; the plus button opens extra chats for parallel work, each able to take on separate jobs. On the training-community side the speaker promotes a 3,400-member AI profit group with a one-hour Harness course, and summarizing that promo section in a single sentence is enough.
Key moments
AI commentary
"What excites me most is the everything-is-a-plugin design; even the interface itself can be swapped. Scheduled tasks are a genuine force multiplier for small teams. Provided the security notes are taken seriously, I see a setup well worth trying on the free tier."
AI assessment
The strongest counterargument gathers where desktop agents are tested on the freedom-versus-risk balance. Critics argue that making every layer replaceable complicates auditing, that third-party plugins carry supply-chain risk, and that a 60 percent adoption rate magnifies that exposure. I find this criticism far from baseless; when installing a plugin, maintenance status deserves as much attention as sources and star counts.
The list of gaps is not short either: the release is still a preview, the interface shifts between versions, breaking compatibility changes are openly announced, and Linux uses the browser route instead of a native app. Schedules not firing while the app is closed, mobile access still sitting on the roadmap, and unclear enterprise audit trails draw limits for daily operations. This picture makes backup and rollback plans mandatory before connecting critical work.
The speaker's possible interest also belongs in the frame; the closing section is devoted to promoting a training community and courses, so part of the praise should be read through that lens. Still, desktop convenience, free model options and the scheduler together produce a clear practical takeaway for small teams and solo makers: automate low-risk, recurring, well-defined jobs, keep private data outside, and keep the plugin list short and familiar. I consider it worth trying within that frame, but not with production data.
Sources
7 links; no other published story cites them. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
deepseek · ai-agent · desktop-app · plugins · automation · open-source