DeepSeek Harness Ships as a Bare WebUI — the Community Built the IDE Around It

DeepSeek Harness has been released, and the project is also open-sourced.

This time, DeepSeek Harness is built based on the Cordis plugin system, allowing you to extend plugins arbitrarily.

This article will talk about the plugins DeepSeek has provided.

Yes, these are the plugins created by the folks who made those anime avatars during the DeepSeek Harness closed beta. There are many useless plugins, like the whale girl and the whale desktop pet, but some are quite good. This article will discuss them.

However, DeepSeek Harness being translated as DSH Bachelor is just too much.

0x00 Development Experience/UI

dsh-better-sidebar

Since DeepSeek starts as a WebUI, which is too plain, someone developed DSH-better-sidebar: a sidebar workspace for file viewing/editing, terminal, Git, sub-agents, and extensible tabs. It looks just like Codex.

GitHub: https://github.com/omdsh-dev/DSH-better-sidebar

After installation, it looks like this.

Damn, it's practically identical to Codex.

dsh-at-file

Another feature I've seen in Codex, this plugin supports referencing files using the @ method.

Usage in Codex:

dsh-at-file GitHub: https://github.com/omdsh-dev/dsh-at-file

Screenshot after installing dsh-at-file is below; you can now use @ to reference files for editing and modification.

dsh-TUI

This is probably the most popular plugin in DSH. Since DSH officially hasn't released any CLI or TUI form, TUI can only be extended through plugins.

GitHub: https://github.com/ccch1mneyyy/dsh-TUI

After installation, execute dsh --profile cc-tui in the corresponding .bin directory to enter DeepSeek Harness TUI.

I checked, and basically all commonly used commands are covered.

This plugin and better-sidebar go in different directions.

better-sidebar adds a workspace to the WebUI, while dsh-TUI moves the entire interaction back to the terminal. If you're used to browsing file trees and previewing Markdown in the browser, install better-sidebar; if you're already inseparable from the Claude Code or Codex CLI style, install dsh-TUI.

deepseek-harness-desktop

This is arguably the hottest DeepSeek Harness plugin recently.

This plugin does what DeepSeek Harness didn't do.

It essentially creates an Electron desktop client for DeepSeek, similar to the Codex App. It automatically starts and manages the local Harness service, integrates a system tray + desktop window, no need to install Node.js, no need to type commands.

GitHub: https://github.com/anywhere-labs/deepseek-harness-desktop

Hmm, it looks the part now.

But it should be noted that deepseek-harness-desktop is a community project, not the official DeepSeek desktop client. It currently mainly supports macOS and Windows; a plugin marketplace and phone remote access are still in the planning stages.

With these plugins, DeepSeek Harness is gradually becoming like Codex.

So why did DeepSeek Harness release as a WebUI? What's good about this form? It's neither appealing as a CLI (for geek programmers) nor as an app for non-programmers...

Based on the plugin system concept, interactive TUI and app forms can only be filled in by the community.

So I can probably understand Harness's approach: they treat TUI and app as a kind of visual plugin.

I suddenly realized this might actually be the right thing to do.

Besides the TUI and app plugins above, the community has also provided two other plugins; install according to your preferences.

oh-dsh: TUI/Desktop/Web three-in-one community distribution

GitHub: https://github.com/hust-open-atom-club/oh-dsh

dsh-tianshu-tui: Interactive terminal UI plugin on the official DeepSeek Harness

GitHub: https://github.com/huiliyi37/dsh-tianshu-tui

The Tianshu suite is heavier. Besides TUI, it also adds TDD workflows, evidence gates, visual bridges, code retrieval, memory, and cross-session recall.

0x01 Vision / Multimodal

It has been clearly stated before that multimodal is important, but it's not DeepSeek's main focus, and I think the image above fits here too.

So the vision/multimodal part has again been handed over to the community.

dsh-vision-toolkit

This is an image recognition ToolKit capable of image Q&A, long screenshot OCR, UI restoration, and pixel diff.

GitHub: https://github.com/Anionex/dsh-vision-toolkit

However, note that installation requires an external vision model API key to read images, because the DeepSeek API cannot read images.

I configured an API key, and now I can happily read images.

I asked DeepSeek to read the image above, and sure enough, it didn't recognize Liangzi/Saint/Xiao Nanliang.

dsh-vision-proxy

If you don't want to configure an external API, but you have Ollama locally with a vision model, you can use dsh-vision-proxy.

GitHub: https://github.com/Flyvhidbwo/dsh-vision-proxy

Its autoLocalOllama is enabled by default, probing http://localhost:11434 at startup. If detected, it places the local Ollama at the front of the fallback chain.

The benefit is straightforward: screenshots don't need to leave your machine. If a project has backend addresses, intranet data, or client pages, I'd prefer this approach.

The main differences between it and dsh-vision-toolkit are as follows.

0x02 Memory / Notifications / Security

dsh-mnemon

dsh-mnemon is a local three-tier memory system. It organizes the hot memory needed for each turn, the project files that need full reading, and the long-term memory recalled on demand, all in the same workspace. Other Agents that also connect to Mnemon and use the same accessible local Mnemon storage can share long-term memory with DSH.

GitHub: https://github.com/omdsh-dev/dsh-mnemon

If you have a particular preference for memory, you can install this.

Note that Mnemon doesn't magically have memory just by installing the plugin; it also requires a local Mnemon service. Its advantage is that memory is stored locally in SQLite, JSON, and Markdown, and it allows multiple Agents connected to Mnemon to share it.

Simply put, dsh-mnemon is suitable for those who seriously maintain long-term memory. If you just want to reduce repetitive explanations in the current project, DSH's built-in project description is sufficient.

distill

This is also a memory-oriented plugin, but its advantage lies in background subagent reflection + automatically distilling experience into skills.

GitHub: https://github.com/LoserFox/distill

dsh-automation

Since DSH lacks scheduled tasks, the community created a scheduled task plugin. It schedules coding tasks to run in new sessions, preserving an auditable history.

GitHub: https://github.com/titanwings/dsh-automation

Each task runs in a new Session, not inheriting the original session's approvals and context, only allowing read-only or workspace write permissions. Operations requiring manual approval will simply fail, rather than getting stuck waiting for you to confirm in the middle of the night.

Besides scheduled tasks, the community has other fun things, like the one below.

dsh-workflow

This is a Codex/Claude Code-style UltraCode multi-Agent scheduling system, a Workflow layer that can generate/save/govern/recover.

GitHub: https://github.com/icetomoyo/dsh\_workflow

DSH's original multi-Agent was more like a one-time dispatch, calling a few people over to finish a task and that's it.

If you only occasionally open two subagents, there's no need to install this. For tasks like code review, parallel research, and fixed delivery processes that are often repeated, this plugin becomes valuable.

0x03 Toolset

dsh-data-agent

A Data Agent that connects to databases, allowing you to visually add database operations in DeepSeek Harness.

GitHub: https://github.com/omdsh-dev/dsh-data-agent

It currently supports MySQL, PostgreSQL, SQLite, Oracle, Hive, and Impala. After configuring the connection, the Agent can directly view table structures, write SQL, and execute queries.

However, I suggest caution with this plugin. It can not only query data but also execute write operations, and each call auto-commits.

dsh-openpencil

A plugin that lets an Agent drive a real, editable, interactive design canvas, rather than returning a generated image.

GitHub: https://github.com/ZSeven-W/dsh-openpencil

dsh-market

This is a plugin marketplace for DSH, where you can basically find all DeepSeek Harness plugins.

It looks just like the VS Code plugin marketplace.

GitHub: https://github.com/dsh-market/dsh-market

If you plan to use DeepSeek Harness long-term, this might be the first one to install. The marketplace already includes over 300 plugins, and you can search, install, update, and uninstall all within the interface.

0x04 Skins / Desktop Pets

Plugins that should be recommended must be recommended, and those that should be criticized must be criticized.

I criticized DeepSeek Harness before, and now I'll criticize it again.

Did you, DeepSeek Harness, recruit so many ACM gold medalists just to make skins?

So the Harness lead asked people to leave their GitHub IDs on a well-known social platform and even asked for DMs, and the result was all these skin plugins?

Do you need ACM gold medalists for skins? Such a waste of talent? Or did one of the thousands of issues you raised ask everyone to make a skin?

Isn't this the common saying: interview for rocket science, work on tightening screws.

So I came up with a matching couplet: closed beta finds gold medalists, testing ends with a whale girl.

And that's the plugin to recommend next, ACM Whale Girl

Ah, not that.

dsh-deep-whale

Come on, check out the Deep Sea Maid Workshop skin.

GitHub: https://github.com/Small-tailqwq/dsh-deep-whale

It can provide you with extremely strong emotional value.

whale-girl

Oh right, here's a Deep Sea Maid desktop pet for you, so you don't have to work anymore and can just enjoy a laid-back life.

GitHub: https://github.com/vlln/whale-girl

The most fitting comment I've seen recently about DeepSeek Harness is:

Between 'so awesome' and 'so fun,' they chose 'so fun.'

My current judgment on DeepSeek Harness: when it was released, it indeed only provided a skeleton, and Cordis completely handed the extension rights over to the community.

In just a few days, TUI, desktop client, vision, memory, and workflows have all been completed by the community.

But now, the most important thing for DeepSeek Harness is whether it can support developers worldwide in building the ecosystem together, just like Android, iOS, and VS Code. I think what really needs to be observed is whether Cordis's interfaces can remain stable long-term, whether there are unified standards for plugin installation, updates, permissions, and compatibility; whether the development documentation, scaffolding, debugging tools, and distribution channels can allow ordinary developers to quickly create a plugin.

If this AI Infra is done well, then DeepSeek officially only needs to continue doing a good job on the model, Agent Runtime, and Cordis. The desktop client, TUI, vision, workflows, and even various industry Agents can all be left to the community. Then DeepSeek Harness could truly become the VS Code of the Agent era: the official provides a stable kernel, and developers decide how far its extensibility goes.

If it can't be done, then the final outcome might be very bleak.