DeepSeek Harness turns plugins into an agent workspace

DeepSeek Harness is turning the agent plugin model from an app-store metaphor into a workspace assembly problem. Woshipm reported that developer preview version 0.1 went online on the evening of August 13, then passed 100000 GitHub stars in 3 days and had more than 4300 third-party open-source plugins after the same span. The peak was 99 plugins per hour. That is the obvious growth story, but it is not the useful one.

The sharper signal is duplication. In the desktop track, woshipm counted 55 projects with the same name, dsh-desktop; among 33 projects named deepseek-harness-desktop, the top-ranked one had 6897 stars while the remaining 32 had 255 stars combined.

That pattern makes Harness less like a single product to install and more like an environment to inspect. The question shifts from which agent wins on default behavior to which pieces of a daily toolchain should become replaceable, scriptable, and visible enough for an agent to operate inside them.

The plugin count is real, but it is not the product

DeepSeek Harness can look finished if the only number on screen is the plugin count. According to woshipm, it had more than 4300 third-party open-source plugins 3 days after release, and surpassed 100000 GitHub stars in 3 days. Its developer preview version 0.1 went online on the evening of August 13. Plugin growth peaked at 99 plugins per hour.

That is ecosystem motion, but not yet product maturity.

The same ledger also describes a sandbox still forming its edges. Woshipm says DeepSeek Harness is in internal testing, has no TUI and no client, and can only be opened through a web interface. Ifanr says the official installation paths are source installation and quick startup through Node.js. It also notes that launching with `npx @deepseek-ai/dsh web` ties the service to the terminal process: close that terminal or end the process, and the service stops.

The repository data is noisy too. In the desktop track, woshipm counted 55 projects with the same name, `dsh-desktop`. Among 33 projects named `deepseek-harness-desktop`, the top-ranked one had 6897 stars, while the remaining 32 had 255 stars combined. Woshipm also identified 216 older tag-riding projects that used the `dsh-plugin` tag while being unrelated to DeepSeek Harness; those projects had 47105 stars, accounting for 23%.

Most plugins were small bursts, not hardened tools. Woshipm found that only 87 of the 4300 plugins had more than 100 stars, accounting for 2%. More than half had a production time of no more than 6 hours, from repository creation to last commit, and the median active span was about 3.58 hours. Ifanr adds that Anywhere Lab built a third-party desktop client with 14000 GitHub Stars and a plugin marketplace entry, which shows demand. It also shows the official shape is still incomplete.

From plugin-based agent apps to agent-ready operating environments

Everything is a plugin changes what users are trusting

DeepSeek Harness asks users to trust a stack, not a single extension. According to woshipm, its core design is "everything is a plugin," with a detachable system built around the Cordis kernel. That is broader than a marketplace idea. The software interface, MCP calls, Skills system, sandbox environment, and system prompts can all be treated as hot-swappable parts.

The trust boundary moves inward.

Woshipm describes model adapters, tool registries, session logs, and the Agent Loop itself as plugins. The same design includes a tool system for executing commands and file operations, plus session records for restoring, forking, and replaying context. That means a plugin can sit close to the places where an agent decides what to ask the model, what tool to call, and what previous context to carry forward.

Permission policy becomes the practical brake. DeepSeek Harness includes a policy layer that blocks before high-risk operations, and ifanr says it pops up notifications when it needs a user decision or authorization approval. Those prompts matter because the tool system is not decorative. It is the path toward local files and command execution.

Inspection has to happen before convenience wins. Woshipm notes that `dsh --profile web --dump-config` can show which plugins DeepSeek Harness actually starts. That command is less exciting than a new agent skill, but it answers the first operational question: what is loaded, and therefore what could influence behavior?

The official split between `web` and `headless` also changes the risk shape. Woshipm says `web` is a complete application with a browser interface, while `headless` has no server and receives one task before returning a result. Ifanr also points to a trajectory page showing tool calls, assistant replies, context injection, and Token usage. For practitioners, that page is where trust becomes inspectable.

Installation is getting easier than evaluation

DeepSeek Harness lowers the first barrier in a very familiar way. A command line, then a browser tab. According to ifanr and woshipm, the quick path is `npx @deepseek-ai/dsh web`, and ifanr says the service normally opens at `http://127.0.0.1:3080`. The official routes are installing from source or starting through Node.js; the source route begins with `git clone https://github.com/deepseek-ai/deepseek-harness`.

That matters because the product does not yet ask beginners to choose a full client stack before seeing the interface.

It is still a web-first tool. Woshipm notes that DeepSeek Harness currently has no TUI and no client, and can only be opened through a web interface.

Even that setup step is being delegated. Ifanr says Codex can be given the DeepSeek Harness GitHub repository link and full access permissions to install it locally. The same outlet says WorkBuddy can do the local deployment when prompted with the repository and the target address `http://127.0.0.1:3080`. This is the practical direction: installation becomes another agent task, not a ritual the user must memorize.

Skills make the entry point feel even lighter. Ifanr gives the example of a PPT Skill, ppt-master, installed by sending the Skill link `https://github.com/hugohe3/ppt-master/blob/main/skills/ppt-master/SKILL.md` in the chat box. Anywhere Lab's third-party desktop client also adds a plugin marketplace entry compared with the official WebUI. But easier discovery is not the same as proof of usefulness.

The evaluation problem shows up in the visible add-ons. Ifanr counted 189 skin plugins in the DSH skin market, including Deep Sea Maid Workshop, Back to 2005 advertising portal website, ocean skin, sliding ancestor transformation, and sticky notes. Those entries may make the workspace feel alive, but they do not by themselves prove better agent loops, safer tool execution, or cleaner context handling.

Woshipm's author frames the stronger fit more narrowly: developers studying Agent Loop mechanics, teams building Agent products without starting from zero, and heavy users who want to replace models, tools, sandboxes, interfaces, and permission policies themselves. Installation is becoming casual. Evaluation is becoming the actual work.

The useful lesson today is observability, not replacement

DeepSeek Harness is most useful today as a way to watch an agent work, not as an automatic replacement for Claude Code, Codex, Cursor, or a stable IDE. That distinction matters because the project itself was developed with Claude Code, according to woshipm, and ifanr describes Codex and WorkBuddy as tools that can install it locally from the GitHub repository or deploy it to `http://127.0.0.1:3080`. The point is not that the harness makes those tools obsolete. It shows which parts of an agent session should be visible.

The valuable surface is the trace.

ifanr says DeepSeek Harness can split a user request into multiple trackable tasks, then use goal mode to keep those tasks tied to the original requirement. It also has permission notifications when a decision or authorization is needed. Its trajectory page exposes tool calls, assistant replies, context injection, and Token usage, which turns an agent run from a black-box chat into something closer to an inspectable build log.

That is why model switching matters less as a leaderboard claim and more as an evaluation tool. woshipm notes that DeepSeek Harness supports adding third-party large models in settings, and its Agent presets include standard mode and PTC mode, where the Code Mode SDK lets the model combine multi-step operations with TypeScript programs. The official forms also separate usage patterns: `web` is a browser application, while `headless` takes one task and returns a result. Those choices let teams compare behavior across model, interface, and execution style.

The cost evidence is still narrow. woshipm cites one run with DeepSeek-V4-Flash: 1 round, 14 steps, LLM 3m55s, tool calls 18.2s, average first token 1.2s, 112 tok/s, cache hit 93%, input 680K tok, and output 24.5K tok. Useful data, but not a budget model.

The same caution applies to multimodal work. The 0820 update log added native image-request configuration for the DeepSeek adapter, image-text input for `/goal` and `/plan`, and file or session references through the `@` menu. ifanr also notes that DeepSeek-V4-Pro is not multimodal, so Harness can send images only to models that support vision. Before that update, image-heavy multi-turn chats could crash from Token overrun or unsupported API formats. Observability is the practical lesson because it makes those failure modes findable.

When to assemble the agent environment, and when to leave it alone

• You want to study or modify how an agent actually works rather than just use a finished coding assistant. Use DeepSeek Harness. Its design exposes the Agent Loop, model adapters, tool registries, session logs, sandbox behavior, system prompts, and permission policy as replaceable plugin components, and `dsh --profile web --dump-config` can show which plugins actually start.
• Your team is building an agent product and does not want to start from zero. Use DeepSeek Harness as a base layer. The source material describes it as more suitable for teams building their own Agent products, with official `web` and `headless` forms, a Cordis-driven plugin system, session records for restoring, forking, and replaying context, and a tool system for commands and file operations.
• You want a local browser-based agent workspace with inspectable task progress and approval gates. Try DeepSeek Harness through `npx @deepseek-ai/dsh web`. The sources say it normally opens at `http://127.0.0.1:3080`, can split a request into multiple trackable tasks, shows a trajectory page with tool calls, assistant replies, context injection, and Token usage, and asks for user decisions or authorization before certain operations.
• You need a polished desktop client or terminal UI today. Do not assume the official DeepSeek Harness provides that. The sources say DeepSeek Harness currently has no TUI and no client, and can only be opened in a web interface. A third-party Anywhere Lab desktop client exists and adds a plugin marketplace entry, but it is not described as the official client.
• You want the agent environment to include the desktop operating system itself. Watch Omarchy rather than DeepSeek Harness. Omarchy is a Linux distribution; Omarchy 4.0 rewrote the desktop environment into one process called Quickshell, put the status bar, launcher, menus, notifications, and lock screen inside it, changed system configuration into Lua scripts, and lets external programs directly send commands to the shell.
• You are evaluating plugin ecosystems by raw plugin count or GitHub stars. Be cautious. DeepSeek Harness had more than 4300 third-party open-source plugins 3 days after release and surpassed 100000 GitHub stars in 3 days, but only 87 projects among 4300 had more than 100 stars, 216 older tag-riding projects were identified as unrelated, and more than half of plugins had a production time of no more than 6 hours.

Omarchy shows where the harness may go next

DeepSeek Harness stops at the application boundary. Its "everything is a plugin" design, described by woshipm, makes model adapters, tool registries, session logs, and the Agent Loop replaceable inside the agent shell. Omarchy points to the next boundary: the desktop itself. According to geekpark, Omarchy is a Linux distribution, and version 4.0 released on August 14 with AI installed directly into the operating system.

That matters because Omarchy 4.0 does not only add an assistant window. Geekpark says the whole desktop environment was rewritten into one process called Quickshell, with the status bar, launcher, menus, notifications, and lock screen all inside it. System configuration moved into Lua scripts. A plugin is a Git repository.

External programs can send commands directly to the shell.

This is the same harness idea, lowered one layer, and the provider choice makes the direction clearer. During boot setup in Omarchy 4.0, users can choose one of nine providers as the default programming Agent, including Claude Code, Codex, Gemini, Grok, and Copilot. The agent is no longer just a product picked after login. It becomes part of how the machine is assembled.

The plugin numbers show why this matters, but also why caution belongs here. Geekpark reported more than 500 plugins one week after Omarchy 4.0, then 1,074 plugins within ten days. Omarchy held its first plugin competition on August 19. That pace resembles the DeepSeek Harness pattern: once the extension point is simple enough, users stop waiting for the vendor roadmap.

For now, this remains mostly a Linux path. Omarchy 4.0 reduced the system image to under 6GB and made installation 30% faster, according to geekpark, but its deeper bet is not installation speed. It is an AI-readable desktop: Lua for configuration, Git repositories for plugins, Quickshell for visible surfaces, and shell commands for action.

Treat DeepSeek Harness as a sandbox before it becomes a daily driver. Run it with `npx @deepseek-ai/dsh web`, and remember that closing the terminal or ending the process stops the service; according to ifanr, npm uninstall is not required.

In your own work, watch the trajectory page first. It exposes tool calls, assistant replies, context injection, and Token usage, which makes it a good place to see where an agent is guessing, carrying stale context, or spending budget. The 0820 update also matters: `/goal`, `/plan`, and the `@` menu can accept image-text input when the chosen model supports vision. DeepSeek-V4-Pro does not.

Audit the boring parts next. Local folders, repeatable commands, text config, and plugin-like components are the places where an agent can act safely. Cordis handles plugin loading, unloading, and dependencies; the skin market's 189 plugins show how quickly the surface area can grow.

For readers outside China

• Availability: DeepSeek Harness is open-source enough to be installed from source with `git clone https://github.com/deepseek-ai/deepseek-harness` or launched through Node.js with `npx @deepseek-ai/dsh web`, but the sources also say it is currently in internal testing. The official forms are `web`, a browser interface, and `headless`, which receives one task before returning a result. The source material does not say whether it is officially available outside China, whether hosted services are region-limited, or whether support is offered in English. Omarchy is described as a Linux distribution, but availability outside China is not disclosed in sources.
• Pricing: Pricing is not disclosed in sources for DeepSeek Harness, the Anywhere Lab desktop client, Omarchy, or Omarchy plugins. The only funding figure given is that the Omacom Foundation's first $8 million came from eight people, with each person contributing $1 million; that is foundation funding, not product pricing.
• Closest Western equivalents: Claude Code, Codex, Gemini, Grok, and Copilot are named as programming Agent provider choices in Omarchy 4.0.; DeepSeek Harness is closest to an inspectable agent runtime or agent workbench rather than a single coding assistant: it exposes model adapters, tool registries, session logs, the Agent Loop, MCP calls, Skills, sandbox environment, and prompts as plugins.; Omarchy is closer to an AI-native Linux desktop environment than to a plugin inside an editor, because its shell, configuration, and desktop components are made scriptable and commandable.
• Data residency: The sources do not disclose data residency, hosting location, telemetry behavior, retention policy, or whether prompts and files are sent to servers outside the user's machine. DeepSeek Harness can run locally in a browser at `http://127.0.0.1:3080`, supports adding third-party large models in settings, and can send images to models that support vision after the multimodal update, but the source material does not cover where model requests are processed or stored.

Sources

• woshipm DeepSeek Harness 上线 3 天，4300 个第三方插件都在干什么？ https://woshipm.com/ai/6447872.html
• woshipm DeepSeek Harness 安装，初体验，没有惊喜。 https://woshipm.com/share/6447430.html
• ifanr 零基础速通DeepSeek Harness，带你玩转赛博乐高 https://ifanr.com/1675781
• geekpark 硅谷押注的下一个 Harness，是整个桌面操作系统 https://geekpark.net/news/369298