DeepSeek Harness v0.2 Brings Official Desktop Apps to Its Open-Source Agent Harness

Michal Sutter at MarkTechPost reports that DeepSeek shipped official desktop apps for DeepSeek Harness (dsh) with the v0.2 preview — installers for Apple silicon macOS and 64-bit Windows, plus npx @deepseek-ai/dsh web. A harness is the runtime that turns a model into an agent: files, commands, plans.

The preview adds a plugin manager, file and diff review in a sidebar, document/spreadsheet/PDF workflows, and Automation Task for scheduled recurring prompts. The desktop build bundles the dsh command so users don’t need a separate Node or pnpm install. Sign-in is DeepSeek account or API key; account models get web search without an extra key.

DeepSeek open-sourced dsh under MIT in August 2026. MarkTechPost says the repo has passed 240,000 GitHub stars and 29,000 forks, and that the harness is not locked to DeepSeek models — OpenAI-compatible endpoints are supported.

We’re watching agent harnesses become the real product surface. Models swap; the loop that reads your files and runs your tools is what people keep open all day. DeepSeek’s bet — everything is a plugin, including the agent loop — is the open-source answer to Claude Code and Codex. Creator mode even lets the agent write and install its own plugins. That’s powerful and a little reckless in the same breath.

Useful? Absolutely, if you’re a developer who wants an MIT harness that can aim at any OpenAI-compatible endpoint. The preview honesty matters: DeepSeek warns that compatibility-breaking changes will follow, and an experimental Claude Code Mods layer is framed as a test, not a promise. We’re for shipping early when the license stays open and the architecture stays replaceable.

The downside if this scales carelessly is the same agent problem we keep naming: a desktop app with file and command access plus scheduled automation is a malware-shaped product if permissions and review UX are weak. Diff review in the sidebar is the right instinct. Scheduled tasks that fire overnight without a human in the loop are where accidents get expensive.

I’m also watching whether “everything is a plugin” becomes a healthy ecosystem or a compatibility tax. 240k stars is demand. Sustaining that without turning every upgrade into a migration is the harder product job.

MarkTechPost compares dsh to Claude Code and OpenAI Codex on license, surfaces and extension models. DeepSeek also uses the harness as an evaluation rig — Code Agent benchmarks ran on “minimal mode,” with DeepSeek-V4-Flash-0731 scoring 82.7 on Terminal Bench 2.1 in that setup.