DeepSeek shipped Harness v0.2 as an official desktop preview. MarkTechPost, writing on 3 October from the release, says installers cover macOS on Apple silicon and 64-bit Windows. You can also run the web UI with `npx @deepseek-ai/dsh web`. DeepSeek’s own account posted the desktop preview around 30 September. The license is MIT, and DeepSeek warns that breaking changes will follow.
Previously, Harness was a terminal preview you started with npx, and you needed your own Node setup. The desktop build bundles the runtime and can install the `dsh` command from the app, so plugin installation no longer needs a separate toolchain. Sign-in uses a DeepSeek account or an API key. Reporting on the release says sessions can use other models through OpenAI-compatible endpoints, and that the default in-app model is DeepSeek-V4.1-Flash.
Own the loop if you want. Do not put client data in it yet.
What actually happened
The product change is distribution. A preview agent runtime now installs like a desktop app: sign in, pick a folder, hand it a task. v0.2 adds a plugin page, so you can install an npm package without living in the terminal, plus an Automation Task plugin for recurring prompts. Creator mode, described in coverage of the release, lets the agent write and install its own plugins. DeepSeek also removed some runtime invariant plugins, which is a breaking change for extensions built on the first preview.
Coverage disagrees on Intel Macs. The MarkTechPost read of the official installers says Apple silicon and Windows x64. Other write-ups mention Intel builds. Use the file DeepSeek actually offers on the download page, not a secondary compatibility list.
The harness is free. The model is not, unless your DeepSeek account still has credit. Pointing it at another host means you pay that host.
Why a small team cares
This is for a technical solo who already runs an agent against a repo and is tired of paying a per-seat subscription for the shell around the model. The Monday test is a spare folder, not a client project. Ask it to sort a CSV or sketch a small script. Then swap the endpoint and see whether the loop still holds.
A designer-owner who does not read diffs should not install this just to “have an agent.” Preview status, plugin breakage, and a model route you configure yourself are the product. Client contracts, credentials, and production repos should stay out until the preview label comes off.
Hype vs useful
Useful: official desktop installers, an MIT license, plugin-shaped architecture, OpenAI-compatible providers in release coverage, and a bundled CLI. That is a real change if you wanted an agent loop you can point away from one vendor.
Not useful: treating v0.2 as a Claude Code replacement. It is a developer preview. DeepSeek has said compatibility-breaking changes are coming. Star counts and “free agent” headlines do not make it safe for client files. One toy folder is enough for the experiment.