DeepSeek Harness (dsh) is DeepSeek's open-source agent harness, released in developer preview under MIT and built on the Cordis meta-framework. Models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and UI are all implemented as plugins, so a harness is composed rather than configured. `npx @deepseek-ai/dsh web` starts a local Web UI. Developer preview: expect compatibility-breaking changes.
What it is
DeepSeek Harness (dsh) is DeepSeek’s own agent harness, released into developer preview in September 2026 under MIT. It reached 218,000 GitHub stars within days of publication, which says less about the code and more about how many people are waiting for a first-party harness from a major open-weight lab. The codebase is TypeScript (pnpm workspace, tsdown, vitest) with a Python side, builds for native, web, and desktop targets, and is iterating fast - the README warns of compatibility-breaking changes ahead.
Everything is a plugin
The design bet is the Cordis meta-framework, described in the paper “A Programming Paradigm for Spatiotemporal Composability.” Models, tools, skills, sessions, sandboxes, filesystems, agent loops, orchestration, and the UI are each implemented as plugins. Nothing is baked in: a harness is composed from parts rather than configured through options. Plugins carry the dsh-plugin GitHub topic, so the ecosystem is meant to be browsable on GitHub the way an app store is browsable.
The practical consequence for model choice: the model layer is a plugin, not a constant. DeepSeek models are obviously first-class, but nothing in the architecture requires them - the same harness runs against whatever provider plugin you install, and local endpoints are just another plugin away.
Install and run
npx @deepseek-ai/dsh web (Node.js) starts a Web UI at 127.0.0.1:3080 and opens it in your browser - use --no-open to skip the launch, and note the SSH-session case only prints the host URL since port forwarding is handled client-side. From source: clone the repo, pnpm install && pnpm run build, then pnpm dsh web. Documentation lives at deepseek-harness.github.io, with a user guide for the Web UI under docs/user/guide, a SAFETY.md notice, and GitHub Discussions plus a Discord for feedback.
Where it fits for local AI
A major open-weight lab shipping its own harness is the missing piece of the local-agent stack falling into place: open models (DeepSeek V4.1 Flash, GLM 5.x, Qwen) plus a plugin-composable harness plus your own hardware is a complete, vendor-independent agent setup. The sandbox, session, and filesystem layers being plugins also means the trust boundary is inspectable - you can read what the agent’s loop and tools actually do before you let it near a machine. Treat the developer-preview label honestly: this is a watch-and-try project right now, not a production dependency.
Sources: GitHub repo · announcement thread · docs