The most useful file in DeepSeek Harness may be SAFETY.md. It says the agent runtime has not had a security audit and must not be treated as secure or production-ready. That is worth knowing before you type the one-line install.
DeepSeek AI published the harness, whose CLI is called dsh, at deepseek-ai/deepseek-harness under the MIT License. Its motto is "Everything is a Plugin". The runtime sits on the Cordis plugin framework, and you extend it by writing plugins.
Official Risk Notes
dsh can run model-generated code and commands, load third-party plugins, and reach whatever network, processes, credentials and files you give it. Sandboxing and approval prompts lower the risk without guaranteeing isolation.
The safety file's advice:
- run with least privilege, ideally in a disposable VM or container;
- back up any file it can reach;
- keep sensitive credentials away from it unless you accept the risk;
- read plugins and proposed commands before letting them run.
Starting the Web UI
With Node.js installed:
npx @deepseek-ai/dsh web
A Web UI comes up at http://127.0.0.1:3080. On a local launch it opens your browser; over SSH it just prints the URL, and --no-open skips the browser either way. To build from source, the README runs pnpm install, pnpm run build, then pnpm dsh web. Full docs live at deepseek-harness.github.io/deepseek-harness.
Developer Preview Status
| Item | Official detail |
|---|---|
| Developer | DeepSeek AI |
| License | MIT |
| Main language | TypeScript |
| Architecture | Everything-is-a-plugin, powered by Cordis |
| Status | Developer preview; the README warns of compatibility-breaking changes |
| Repo created | August 13, 2026 |
Release cadence matches the "preview" label. On npm, @deepseek-ai/dsh tags 0.2.0-rc.2 as latest and 0.2.1-alpha.3 as alpha. The GitHub release dsh-v0.2.1-alpha.3 went out on October 14, 2026 (21:32 Beijing time), following alpha.1 on October 3 and alpha.2 on October 9. Pin a version if you build on it.
For the thinking behind Cordis, the README cites the paper A Programming Paradigm for Spatiotemporal Composability (arXiv 2608.25512). Third-party plugins are supposed to carry the dsh-plugin GitHub topic, which is the easiest way to browse what exists.
GitHub Star Ranking
AI Rank's open-source snapshot of October 15, 2026:
| Metric | Value |
|---|---|
| Stars | 249,547 |
| Forks | 30,005 |
| New stars, last day | +894 |
| New stars, last 7 days | +4,479 (+1.8%) |
| Overall rank | #15 (up from #22) |
| Rank by 7-day star growth | #14 |
| Rank by total stars | #6 |
| Contributors | 44 |
Two months after the repo was created, it ranks sixth by total stars. AI Rank places it under AI Engineering, in the agent-workflow SDK and visual-builder subcategories.
→ Project ranking: deepseek-ai/deepseek-harness on AI Rank — daily GitHub star growth
It is a good fit for developers who want an MIT-licensed, plugin-driven agent runtime they can read end to end, for teams sizing up DeepSeek's own agent tooling inside a sandbox, and for plugin authors happy to chase a moving API. Anyone who needs an audited production platform today should take SAFETY.md at its word.
Compare before you choose: star counts move daily — check the current open-source AI project rankings on AI Rank before you commit.
Or 📱 get the AI Rank app to follow them on your phone.
Prepared with automated editorial assistance and checked against official sources and dated AI Rank data on October 15, 2026. We did not run our own tests for this article.