What Is DeepSeek Harness? A Beginner's Guide to dsh
2026-08-18
What DeepSeek Harness is
DeepSeek Harness — usually shortened to dsh — is an open-source agent harness developed by DeepSeek AI and currently in developer preview. It is the layer that turns a language model into an agent that can read and write files, run commands, use tools, and keep working across sessions. You can start it with one command: npx @deepseek-ai/dsh web, which launches the Web UI on your machine. The source lives at github.com/deepseek-ai/deepseek-harness.
Why an agent harness is not a chatbot
A chat window sends text and returns text. A harness gives the model an environment: access to your workspace, tools it can call, a place to store sessions, and boundaries that decide what it may do. That difference matters in practice. In DeepSeek Harness, the model is one component among many, and the harness coordinates everything around it — the sandbox, the approval policy, the session log, and the agent loop that decides when a task is done.
Everything is a plugin
The core design principle is explicit: everything is a plugin. Models, tools, skills, sessions, sandboxes, storage, scheduling, and the UI are all plugins mounted on a kernel called Cordis. Cordis handles mounting, unmounting, and dependency resolution, and every registration is reversible — when a plugin unloads, what it registered is cleaned up automatically. Because capabilities are plugins, you can swap or extend any of them in configuration without changing the DeepSeek Harness source code.
Profiles and bundles
A profile is a named composition stored under the harness home, and web and headless ship as templates. A bundle is an npm package that contributes a configuration layer, declared by a dsh.bundle field in its package.json. To install a plugin you run dsh plugin --profile web add <package> for an npm package, or dsh plugin --profile web add github:owner/repo to install from a repository. GitHub installs fetch source, so the package's prepare script builds it on your machine — pnpm asks you to allow that build explicitly. Prefer pinned commits (github:owner/repo#sha) and only install sources you trust. See the install guide for the full walkthrough.
Every run is traceable
Everything the model sees is written to an append-only session log: system prompts, reasoning, tool calls and their results, subagent scheduling, and context injections. The Trajectory view lets you inspect a run by source, and you can resume, fork, search, or replay the same event stream. For developers debugging agent behavior, that record is the difference between guessing and replaying.
Where dsh plugins come from
Plugins are published to npm or shared as GitHub repositories, and most carry the dsh-plugin topic so they are discoverable. dshplugins.cc tracks that ecosystem: it indexes plugins that use the topic, records stars and download counts daily, and gives every plugin a page with install commands and a health score built from real signals — license, community adoption, documentation, and distribution. Start from the plugin directory or browse by category.
FAQ
Is DeepSeek Harness free and open source? Yes. The source is public under the deepseek-ai GitHub organization and the project is in developer preview.
Do I need a DeepSeek API key? Not necessarily. The model layer is an adapter plugin, so setup depends on which model provider you configure.
Is it stable? Not yet. DeepSeek describes the preview as still being tested by agent harness developers; core plugins and APIs will keep evolving.
How do I install plugins? Run dsh plugin --profile web add <package> or dsh plugin --profile web add github:owner/repo from a checkout, or follow the install guide.
Conclusion
DeepSeek Harness is a developer-preview agent harness whose entire surface is composed of swappable plugins. If you want to try it, start with npx @deepseek-ai/dsh web, then install a couple of plugins from the directory. The fastest path is the install guide, and the health score on every plugin page gives you a quick, real-data check before you add third-party code to your setup.