DeepSeek’s most interesting launch of the past week was not a model. On 29 September it released DeepSeek Harness v0.2, a preview with ready-to-install desktop apps for macOS and Windows, and on 3 October it followed up with a 0.2.1 alpha. IT之家 (IT Home) and 智东西 (Zhidx) both reported the desktop release; MarkTechPost has the English rundown. A “harness” is the runtime that turns a model into an agent — it reads files, runs commands, keeps a plan and calls tools. DeepSeek is now shipping one to people who have never opened a terminal.
What v0.2 does
According to IT之家’s summary of the official notes, the desktop build comes preloaded with common office and development functions. Users can hand it documents, spreadsheets or PDFs and ask for data analysis, charts or slides, or point it at a code project to change and run. Generated files and code diffs appear in a sidebar for review. A new plugin management page installs plugins by npm package name — no command line — and an experimental “Creator mode” lets the agent write and install its own plugins from a chat request. An optional automation plugin schedules recurring tasks with run history.
Under the hood, MarkTechPost notes the harness is MIT-licensed, built on the Cordis framework where the model adapter, tool registry and agent loop are all plugins, and not locked to DeepSeek models — custom OpenAI-compatible endpoints work too. Per the release notes it cites, users signing in with a DeepSeek account get web search without an extra key. DeepSeek also uses a “minimal mode” of the harness for its own Code Agent benchmarks. The GitHub repository shows more than 240,000 stars.
The 3 October build, per its changelog, adds an experimental compatibility layer for Claude Code Mods (DeepSeek stresses it is testing whether that API is broadly a subset of what its plugins can do, not promising full compatibility), a developer-tools bundle with raw session logs, and a --public-url flag for reverse proxies — plus breaking changes for some plugins.
The numbers and the roadmap
Two disclosures stand out in Zhidx’s report. First, DeepSeek says roughly 60% of Harness users (measured among official API users) already use third-party plugins, and it plans an official plugin marketplace with a more stable plugin API. Second, DeepSeek says that by daily active users and daily sessions, Harness has become the most popular coding agent among users of its official API. Eligible account holders are also being offered ¥6 of trial credit.
Then the roadmap: better sandboxing and security, improved agent teams, personalised long-term memory, automated operation of the browser and other GUI software, remote and mobile use, and session sharing and collaboration. DeepSeek also says Harness will “co-evolve” with its models — optimising for real performance inside the harness during model training.
Why this belongs on a search site
DeepSeek’s own Harness page lists research — “find information, verify facts, and cite sources” — alongside coding and office work. Combine that with account-level web search today and browser automation on the roadmap, and the direction is obvious: an open-source Chinese agent that will increasingly visit, read and act on websites on a user’s behalf, from desktops worldwide.
That has three consequences for site owners. Your pages will be consumed by agents that summarise and cite rather than browse and convert, so extractable facts, clear dates and unambiguous product data matter more. Your bot management needs a view on agent traffic that arrives through ordinary browser automation rather than a declared crawler. And plugins are becoming a distribution channel in their own right — the same shift we flagged when Baidu opened Search Widgets and asked webmasters to submit skills instead of pages.
Our take: DeepSeek has been the efficiency story of Chinese AI — routing Pro traffic to V4.1 Flash because the small model won. Harness is the distribution story. A free, MIT-licensed agent that ships with plugins, scheduling and soon a browser is how a model company gets into everyday work without building a search engine. And DeepSeek is honest that the security story is unfinished — consistent with its warning that agents can’t be trusted unsupervised. Treat it as a preview. But start planning for the visitors it will send.