DeepSeek Harness Plugins, Explained: Install, Add, and Build on the "Everything Is a Plugin" Agent Runtime

The one-line answer: DeepSeek Harness is not a model. It is the open-source agent runtime DeepSeek released on August 13, 2026 under an MIT license, and its entire bet is that every capability an agent needs — model, tools, skills, session state, UI, even the main loop — should be a swappable plugin. The models you run inside it are DeepSeek V4 Flash and DeepSeek V4 Pro, and you do not have to believe a word of the marketing to feel the difference: run npx @deepseek-ai/dsh web, open 127.0.0.1:3080, and in under a minute you have a local coding agent whose terminal, sidebar, web search, and memory are all plugins you can add, remove, or rewrite without touching the framework source. The GitHub repository that shipped this hit 50,000 stars in its first 12 hours and sits at 125,959 stars as of August 16, 2026. This article is the practical plugin guide the page-one results skip: what DSH actually is, what the plugin ecosystem contains right now, how to install and write plugins, and where the preview genuinely falls down.

What DeepSeek Harness actually is (and what it is not)

DeepSeek's internal formula, confirmed on the official Harness page, is Model + Harness = Agent: the model thinks and reasons, the harness gives the agent the ability to understand its environment, use tools, and keep working across a real task. Anthropic popularized the term and Claude Code is the reference implementation; DeepSeek's version is open source, runs entirely on your machine, and is aimed at the same territory as Claude Code and OpenAI Codex. That positioning is why calling DSH "a model" is wrong and why this article is not a model review: the harness is the layer that connects a model to tools, files, a terminal, and a web UI — and in DeepSeek's design, every one of those connections is a plugin.

The release facts, all verified August 16, 2026: the repo deepseek-ai/deepseek-harness went public the evening of August 13, Beijing time, as a v0.1 developer preview under MIT license. It passed 10,000 stars in under two hours and 50,000 within twelve — roughly 80x the pace of OpenClaw, previously the fastest-growing GitHub repository — and later reports put it past 100,000 by the 42-hour mark. The GitHub API today shows 125,959 stars and 12,526 forks. It is not a cloud service: you install Node.js, run npx @deepseek-ai/dsh web, and the Web UI serves on http://127.0.0.1:3080 with your sessions, logs, and data staying local. The version on npm is 0.1.0-rc.6, and the README states the project is in developer preview and "iterating rapidly," with compatibility-breaking changes expected.

"Everything is a plugin" — what the slogan actually means

The framework is built on Cordis, a plugin meta-framework extracted from the Koishi bot ecosystem. Cordis is deliberately dumb: it only loads, unloads, and resolves dependencies between plugins, and carries none of the agent's actual capabilities. Everything above that — the model adapter, tools, skills, sessions, sandbox, storage, the loop that drives the agent, scheduling, and the web UI — is itself a Cordis plugin, and plugins cooperate through Cordis services and events. You select, replace, or extend any capability at the configuration layer, without editing source code.

That includes the model. The model is just a plugin with an OpenAI-compatible endpoint behind it, so nothing about the framework forces you to stay inside DeepSeek's lineup — the "swap the model" step is a config change, not a fork. The same principle is why the four built-in presets are really four plugin bundles:

• Standard — the full coding agent: file editing, Shell, file and web retrieval, Skills, plans, goals, sub-agents, and workflows.

• PTC — everything in Standard plus the Code Mode SDK, which lets the model write one TypeScript program to chain multi-step tool calls instead of issuing them one by one.

• Minimal — exactly two tools (a persistent bash and an str_replace_editor), built for model benchmarking in minimal environments. This is the mode DeepSeek used for the V4 Flash agent-benchmark runs that its API docs reference.

• Creation — a mode for making new presets: Standard's full capability set plus runtime inspection and in-memory plugin experiments, aimed squarely at plugin developers.

The plugin ecosystem, in numbers

Community plugin development started during internal testing — roughly 300 plugins were built in the first days, including a Windows XP retro skin and a meme plugin — and the numbers have climbed since the public release. The count that matters depends on what you count:

• 288 plugin repositories catalogued on launch night by Chinese tech media, per 36Kr's survey.

• 1,000+ repositories carrying the public GitHub topic tag dsh-plugin within the first days; some counts report 700+.

• 1,117 curated plugins across 1,521 monitored ecosystem repositories, totaling 301,295 GitHub stars, in the community-maintained Oh-My-DSH directory as of August 15, 2026.

• 2,600+ repos in the auto-updating awesome-dsh catalog of the whole dsh-plugin topic, and 368 hand-curated entries across 18 categories in the 0xsline awesome-deepseek-harness list.

The taxonomy in those lists shows where the community went first — and it is exactly the surface an agent runtime needs. Web search: dsh-web-search-exa (Exa-backed, with a keyless fallback) and dsh-web-search-pro (multi-engine routing across DeepSeek, Exa, DuckDuckGo, Bing, Jina, GitHub, Bilibili, and YouTube). RSS and news: dsh-news-plugin pulls 10+ Chinese and English feeds into structured items. Office files: dsh-office lets the model edit Office documents with docx/pdf preview in the web client. Chat history import: dsh-chat-import restores full-fidelity conversations from 13 coding agents — Claude Code, OpenAI Codex, ChatGPT, Cursor, Gemini, and others — as resumable DSH sessions. Mobile: dsh-mobile and dsh-mobileweb-adapter put the UI on a phone. Feishu and Lark: dsh-feishu-bot and dsh-feishu-notify push session events into Feishu, and dsh-im-hub is a multi-platform gateway (Feishu WebSocket, WeCom, Telegram). Voice: dsh-voice adds Edge neural TTS read-aloud and speech-to-text, and dsh-voice-webspeech needs no server and no API keys. Around those sit memory plugins (dsh-memento, dsh-memory-evolve, dsh-recall), an official tool suite (dsh-toolkit), IDE-style sidebars (dsh-better-sidebar, the deepseek-harness-workbench-plugin), vision plugins that add sight to text-only models, and even a Tauri desktop wrapper, deepseek-harness-desktop.

How to install DSH and add plugins

The install path is short, and the plugin command is worth getting exactly right because several tutorials mangle it. The canonical form is dsh plugin --profile web add <package> — the --profile web flag matters, because profiles are how DSH decides which plugin bundle a session runs, and the web profile is the one behind the Web UI.

• Install Node.js (reports indicate 22.19 or newer), then start the Web UI: npx @deepseek-ai/dsh web — it serves at http://127.0.0.1:3080.

• Add a plugin by npm package name: dsh plugin --profile web add <package-name>, for example dsh plugin --profile web add @nanmicoder/dsh-agent-teams.

• Add from a GitHub repo: dsh plugin --profile web add "github:owner/repo#ref" (the #ref pins a branch, tag, or commit).

• Without installing dsh globally: npx -p @deepseek-ai/dsh dsh plugin --profile web add <package-name>.

• After installing a bundle plugin, restart dsh web and refresh the page; verify with dsh plugin --profile web list, or use Settings → Plugins in the Web UI.

One subtlety that trips up first-timers: only packages that declare a dsh.bundle.patch field become active profile layers. A plain dependency installs but stays inert — the bundle declaration is what tells DSH the package is a harness plugin and where it patches in. The dsh plugin command forwards package operations to pnpm, so npm packages, GitHub specs, local paths, file: and link: sources are all supported, and re-running the same add command upgrades the plugin.

How to write your own plugin

Because everything is a plugin, writing one is writing a Cordis plugin that declares a DSH bundle — and the community scaffolding makes it a clean starting point. The plugin registry and development tooling (plugin-registry, dsh-plugin-starter, create-dsh-plugin, and the dsh-plugin-check health-checker) are community infrastructure, but the contract is stable enough to describe: your package must declare dsh.bundle.patch to become an active profile layer, it exposes its capabilities as Cordis services and events, and the runtime decides at load time how it patches into the profile you chose.

For local development the loop is: clone and build your plugin, add it with dsh plugin --profile web add "link:$(pwd)", then iterate — client-side changes need a page refresh, host-side changes need a dsh service restart. That local-link workflow is how most of the ecosystem's first 1,000 plugins were built, and it is worth copying the discipline those projects practice: pin versions. The preview iterates fast enough that a plugin written against one rc can break on the next — the community's ModLens vision plugin, for instance, is documented as needing its version locked to work reliably.

Trajectory and MCP — the two features reviewers keep calling out

Two capabilities are worth knowing about before you decide whether DSH is for you, because they are the parts of the design that a "plugin listicle" will not tell you. The first is trajectory. Every run is traceable: the model's entire experience — system prompts, chain-of-thought, every tool call and its result, sub-agent scheduling, and each context injection — is written to an append-only session log, and restore, forking, retrieval, and replay all operate on that same event stream. Developers on Hacker News flagged this as the killer feature, noting that closed agent products hide exactly this data. The second is MCP. DSH ships an official MCP client with a /mcp command and a Settings tab for servers, and the ecosystem has already built on it — dsh-chatgpt-bridge exposes a DSH session to ChatGPT as an MCP server, and dsh-mcp-panel is an observability panel for the client. If your agent stack already leans on MCP servers, DSH can mount them as plugins.

Where the model comes from

DSH is a harness, not a model provider, so the one thing you must supply is a model endpoint — and this is where the native fit is DeepSeek's own pair. DeepSeek V4 Flash (the efficient 284B-total / 13B-active MoE, roughly $0.15 per million input and $0.29 per million output tokens) is the cheap default; DeepSeek V4 Pro (the 1.6T-total / 49B-active flagship, roughly $0.44 / $0.88) is the high-effort engine. Both are hosted on OrcaRouter at DeepSeek's list price with zero markup, reachable through one OpenAI-compatible key — which fits DSH's model plugin cleanly: point the model plugin at the OrcaRouter endpoint, and you can route between Flash and Pro inside the same harness, with automatic failover to a second provider if one stalls. That is a genuine use case, and an honest one: OrcaRouter does not host DSH itself — DSH runs on your machine, and OrcaRouter is the model layer you point it at.

Where it falls down — the honest part

The same energy that produced 125,000 stars in three days is the energy that produced the rough edges. Read these before you install, not after:

• It is a developer preview, and it says so. The README's own warning is "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." The npm version is a release candidate (0.1.0-rc.6), core plugins and base APIs are still iterating, and the launch UI itself shows a notice aimed at Harness developers. Do not build a production dependency on version numbers that will move under you.

• The harness choice costs real tokens. A third-party comparison cited in Chinese press found another open-source harness used roughly 30% as many tokens as DeepSeek Harness on the same DeepSeek V4 Flash task, and a Composio controlled test ran the same V4 Flash across eight harnesses on 30 tasks and got between 14 and 20 completions — the model identical, the results a third apart. The harness is not a neutral wrapper; it decides how efficiently the model works.

• It feels like a framework, not a finished product. Reviewers' own words: the UI is not friendly to people who do not write code, and rough edges are everywhere. There is no official CLI or TUI — the community built dsh-TUI and dsh-tianshu-tui to fill that hole — and the "everything is a plugin" freedom is exactly what makes the defaults feel unfinished.

• It works, but you have to watch it. The hands-on reports — including the one that coined the "50,000 stars overnight" headline — describe an agent that completes multi-step tasks (rebuilding a site in a different design language, fetching GitHub API data and rendering a chart) but needs supervision, with minor format and interaction errors in the results.

• Local-first means you bring the compute. DSH runs on your machine and your model keys. If you wanted a managed agent, this is the wrong thing.

Bottom line

DeepSeek Harness is the fastest-growing open-source project of 2026 not because it is a better model — it is not a model at all — but because it made the agent itself the platform. The plugin ecosystem is the real story: 100+ capabilities in the official bundles, 1,100+ community plugins catalogued three days after launch, and the fastest way to evaluate the whole bet is to run npx @deepseek-ai/dsh web and add a few plugins. For a developer who wants a local, fully inspectable agent runtime and is comfortable pinning versions on a moving preview, DSH is worth installing today — pair it with DeepSeek V4 Flash and DeepSeek V4 Pro through OrcaRouter at DeepSeek's list price and you have the full stack for the cost of tokens alone. For anyone who needs production stability, cannot babysit a preview, or wants a managed agent, it is not ready yet — wait for the compatibility surface to settle. The framework's claim, that the harness rather than the model is where the value accumulates, is the thing to bet on or argue with — and now, thanks to 1,000-plus open plugins, you can test it yourself.