DeepSeek Harness desktop app enters public preview for macOS and Windows as open source
- DeepSeek Harness is in public preview worldwide and open source, shipping a macOS arm64 .dmg plus a web UI you launch from code, after existing as a harness users ran themselves.
- The harness is built on Cordis, an "everything is a plugin" architecture, and Creator mode lets the agent write and install its own plugins: the demo builds, installs, and verifies a floating Pomodoro timer plugin in 5m 24s.
- Bundled plugins include agent teams, auto approval review, scheduled tasks, voice input, terminal, and subagents, with the model picker showing DeepSeek-V4.1-Flash; scheduled tasks take a prompt and a weekly cron-style spec such as Fridays at 17:00 Asia/Shanghai.
- Developer tools expose a trajectory view with a timeline of turns, tool calls, and runtime context for debugging failed calls and task execution.
- Hacker News commenters split between praise for the plugin and session-visualization work and concerns that the binaries come from China, that data goes to China, and that the self-updating plugin system is an attack surface; others answer that it is open source, unlike Claude Code, and can run in server mode.
Hacker News opinions
Installed it on macOS and everything carried over, all settings and workspaces. Only thing missing is cmd+ and cmd- for font size, so I asked DSH for a plugin and it built one, then a bell for finished tasks. Both worked.
You can change the font size in settings, just not with a keyboard shortcut. For some reason they capped it at 17 max.
What's actually better about it than other harnesses though? Calling it a beast tells me nothing. What can it do that the others can't?
This thread feels astroturfy to me. What an ad.
frontierharness.org puts it on the Pareto frontier, but that benchmark is one-shot tasks with one static model as backend. Harness differences really show up in long-running tasks, and I wish there were benchmarks refreshed more often.
No GitHub Copilot in that comparison either, and it supports BYOK and is pretty capable.
The Cordis architecture is the exciting part and the page barely mentions it. In engineering terms, each plugin declares an activate and deactivate function over shared context plus its required and provided services, so you can enable and disable capabilities without restarting the process.
Binaries come from China, your data goes to China, and the self-updating plugin system is vulnerable to a malicious LLM provider. Anyone using this is crazy.
It's open source. Unlike Claude Code.
As a non-American, handing my data to an American company feels just as risky as handing it to a Chinese one.
DSH still in preview but works flawlessly and updates fast. With the official desktop app out, I hope they add remote connection soon.
You can run it in server mode according to the docs. That said, it doesn't work with Bun and running it in a remote docker container is awkward because of its security model.
The original title said "DeepSeek Harness Desktop app for MacOS and Windows". The new one says nothing. We already knew about DeepSeek Harness, this is an installable app worth differentiating.
Sigh, yet another harness. I'd still try it, honestly, just for the session visualization tooling, which I'd like to steal for my own workflow.