What is DeepSeek-Harness? A Complete Introduction | FindHarnessSkip to main contentLog inSign up

All postsimg:first-child]:pt-0 data-[size=sm]:[--card-spacing:--spacing(3)] data-[size=sm]:has-data-[slot=card-footer]:pb-0 *:[img:first-child]:rounded-t-xl *:[img:last-child]:rounded-b-xl px-5 py-8 md:px-10 md:py-12">Guide2026/08/15<br>What is DeepSeek-Harness? A Complete Introduction<br>DeepSeek-Harness (dsh) is DeepSeek's open-source agent harness where every feature — tools, commands, skills, MCP — is a plugin. Here's how it works.<br>DeepSeek-Harness, commonly shortened to dsh , is an open-source agent harness built by DeepSeek AI. Its defining idea is "everything is a plugin": the CLI, the web UI, tool access, slash commands, skills, and even MCP server connections are all implemented as plugins on top of a general-purpose plugin framework called Cordis. There is no separate manifest format for each capability type — one plugin mechanism covers all of them.

dsh is currently in developer preview (version 0.1.0-rc.5 at the time of writing). The project's own README is explicit about what that means: expect compatibility-breaking changes as the framework iterates.

"Everything is a plugin" — what that actually means

Most agent tools split their extensibility into separate systems: one format for tools, another for slash commands, another for skills, another for MCP configuration. dsh collapses all of that into a single concept. A plugin is just a JS/TS module that exports an apply(ctx, config) function, and depending on what it registers on ctx, it can become any of the following:

You want to add...The plugin registers...A tool the model can callctx.tools.register()A slash command (/xxx) for the UIctx.commandsA skill (injected prompt + tool)a prompt section plus a tool, injected on invocationAn MCP server's toolsone plugin per MCP server, discovering and registering its tools via ctx.tools.register()A hook into agent/tool lifecycle eventsa listener on extension points like agent/pre-step or tools/pre-executeA new LLM providera LlmAdapter registered via ctx.llm.registerAdapter()A background/cron jobctx.jobs<br>Notably, dsh ships official bridge plugins — dsh-hooks-claude-code and dsh-hooks-codex — that translate an existing Claude Code or Codex hooks.json into dsh's own hook extension points, so hook configs from those ecosystems can be reused rather than rewritten.

If you've used a harness where skills, commands, and MCP connections each need their own config file, this is the biggest conceptual shift: in dsh, they're all just code that calls ctx.

Installing dsh

The fastest way to try it is to run it straight from npm, no clone required:

npx @deepseek-ai/dsh web

This starts the web UI, listening by default on http://127.0.0.1:3080.

If you want to work from source instead — useful if you're developing your own plugin against the latest framework code — clone and build the repo directly:

git clone https://github.com/deepseek-ai/deepseek-harness.git<br>cd deepseek-harness<br>pnpm install<br>pnpm run build<br>pnpm dsh web

Profiles: how dsh keeps configurations separate

A profile is a named, runnable configuration — a directory under $DSH_HOME/profiles/ (with $DSH_HOME defaulting to ~/.dsh) that lists which plugin bundles are active and in what order. You start dsh against a specific profile:

dsh --profile

dsh web is actually shorthand for --profile web, and dsh --profile headless "" runs a one-shot task with no UI at all. Profiles are how you keep, say, a minimal headless automation setup separate from a plugin-loaded web UI setup, without them stepping on each other's configuration.

Plugins are added to a profile with:

dsh plugin --profile add

We cover this command — and the npm/GitHub/local-path sources it accepts — in detail in How to Install DeepSeek-Harness Plugins.

The plugin ecosystem today

Here's the part that matters if you're trying to find plugins worth installing: DeepSeek AI does not run an official plugin marketplace or registry. The only first-party discovery mechanism is a convention — plugin authors are encouraged to tag their GitHub repository with the topic dsh-plugin so it can be found by search. There's no central review process and no official list.

Into that gap, the community has built its own directories. The most actively maintained one is awesome-dsh-plugin/awesome-dsh-plugin, a curated, hand-reviewed list that currently tracks 365 plugins across 11 categories — UI Enhancements, Tools & Capabilities, Workflow & Automation, Memory, Development & Runtime, and more.

That's exactly the gap FindHarness exists to close. We index those community-discovered plugins with real GitHub metadata — stars, license, last-push date, and full READMEs — so you can browse, search, and compare them like a proper catalog instead of scrolling a single giant README. Start with /plugins to search the full list, or /categories to browse by what a plugin actually does.

Where to go next

Ready to install something:...