DeepSeek Harness Plugins: Install, Market, Uninstall

DeepSeek Harness is the fastest way to run a plugin-based coding agent locally, and its real power is the ecosystem: one command installs skills from GitHub, a community market adds skins and memory plugins, and cleanup is a single folder delete. This guide walks through the three operations that matter — installing a Skill, grabbing plugins from the market, and removing everything cleanly — so you can go from zero to a customized harness in about ten minutes.

Why the plugin layer is the part that counts

DeepSeek Harness (dsh) is DeepSeek's open-source agent runtime built on the Cordis microkernel, and its tagline is "everything is a plugin": the model adapter, tools, skills, sessions, sandbox, storage, scheduler and even the UI you see are all plugins. That architecture turns customization into installation. Instead of forking the repo to change behavior, you add or remove plugins with a command or a chat message, and side effects are tracked so uninstalls are clean.

One-command launch before anything else

You only need Node.js 18+ on the machine. The npx launch starts a web UI on http://127.0.0.1:3080 and opens it in your browser:

npx @deepseek-ai/dsh web

If you prefer a desktop client, the community build by Anywhere Lab (about 14,000 stars on GitHub) wraps the same Web UI and adds a plugin-market entry point. Running from source is also fine, but for plugin experiments the npx route keeps things disposable. See our earlier DeepSeek Harness quickstart for the full first-run walkthrough.

Install a Skill straight from a GitHub link

Skills are the easiest way to extend the harness — they are plain prompt-plus-workflow packages with a SKILL.md, and you do not need to clone anything. Find a skill you want, paste its GitHub URL into the chat, and ask dsh to install it. A community walkthrough demonstrates this with ppt-master, a PowerPoint-generation skill: sending the SKILL.md URL installs it, and a follow-up prompt like "make a deck about this topic" hands the whole design flow to the agent.

# terminal equivalent for plugin installs
dsh plugin --profile web add github:csyangwen/dsh-memory-evolve

Since the 0820 update, the /goal and /plan commands also accept images and file references via the @ menu, so you can attach screenshots directly into skill-driven tasks — as long as the model adapter supports vision (see the note below).

Shop the plugin and skin market

Beyond skills, the community publishes UI skins and memory plugins. The dsh-skin-market site lists 189 skin plugins — retro portals, deep-sea themes, sticky notes and more — all installable from inside the Web UI. Memory plugins such as dsh-memory-evolve swap the plain chat/trajectory header for a dashboard with skills, todos and persistent memory. There is no central app store to sign up for; you add plugins by GitHub reference and the Cordis kernel handles loading, dependency resolution and reversible unload.

Uninstall: three scenarios, no leftovers

npx launch: nothing was installed globally. Close the terminal or kill the process and the service stops — no npm uninstall needed.
Source clone: stop the service, delete the deepseek-harness folder.
Desktop client: uninstall like any normal app.

To also wipe sessions, config and credentials, delete the ~/.dsh directory. That is the entire cleanup story.

Practice tips

Match the model to the job: DeepSeek-V4-Pro is text-only, so multimodal prompts fail with the default adapter. The rc.8 update wired image input through the harness pipeline; you still need a vision-capable model (for example via the dsh-vision bridge) to actually see. Details in our rc.8 multimodal write-up.
Use creative mode for UI experiments: skin and UI plugins behave better when you switch the harness to creative mode before iterating.
Expect breaking changes: the project is in developer preview; pin plugin versions if you build on top.
Delegate the install: you can hand the repo URL to Codex or WorkBuddy and let an agent install dsh for you — the first step of learning a tool can now be delegating the learning.

FAQ

Do I need a GPU or special hardware to run DeepSeek Harness?

No. The harness itself runs on Node.js 18+ and is hardware-agnostic; model inference happens on DeepSeek's API, not your machine.

Can DeepSeek Harness see images?

Since the 0820/rc.8 update, the input pipeline accepts images in chat, /goal and /plan — but only if the model adapter supports vision. DeepSeek-V4-Pro does not, so pair it with an OpenAI-compatible vision model (for example via the dsh-vision plugin) for image tasks.

How do I remove DeepSeek Harness completely?

Stop the process (an npx launch installs nothing globally), delete the project folder if you cloned the source, and remove ~/.dsh to clear sessions, config and credentials.