Strip Out the 35,000-Plugin Hype: What Can DeepSeek Harness Actually Do That Claude and Cursor Can't?
DeepSeek Harness, or DSH for short, is an open-source coding and agent tool that ships with both web and desktop interfaces. According to a recent post on DeepSeek’s official WeChat account, DSH has reached 1.54 million daily active users and roughly 35,000 third-party plugins. The team also announced plans to build an official plugin marketplace and tighten compatibility and security as part of its pitch to become the operating system of the agent era. Anyone who writes code daily in Claude Code, Cursor, or OpenAI’s Codex might wonder how that math works: the official marketplaces for Claude Code and Cursor each list about 500 plugins, so where does a 70x gap come from?
The more practical question is what you actually gain by spending time tinkering with DSH and its community plugins compared to mainstream coding tools. When DSH first went open-source on August 13, its headline feature was that plugins could swap out even the software’s internal components on the fly, a teardown architecture that felt like overkill for everyday coding at the time. Three months later, a wave of popular community plugins has pulled off several workflows that competing tools simply cannot match.
How That 70x Plugin Gap Is Counted
The answer lies in how DSH counts plugins by default. Every time one of those 1.54 million daily active users talks to DeepSeek’s official backend, the client automatically attaches the list of currently active plugins. Even a private local plugin sitting on one developer’s laptop, never published anywhere, gets counted toward the total as long as it is turned on. The interface also puts several plugin-creation buttons right in front of the user: connect an external tool, set a scheduled reminder, or ask the AI to whip up a desktop widget, and the system uses a built-in template to generate a standalone local plugin package on the spot.
After filtering out broken half-baked projects that fail to install, the community-maintained installable directory indexes 9,714 public plugins. Nearly 60% of them have at most 1 star or bookmark in the open-source community, and cosmetic add-ons like custom themes and desktop pets make up over a quarter of the catalog. Only 271 plugins have earned more than 100 stars, while the top tier sits at several thousand to over 10,000 stars.
Compare that with the official stores next door: both the Claude Code official marketplace and the Cursor official marketplace curate roughly 500 manually reviewed feature bundles each. Stacking an unfiltered count of local micro-widgets against vetted marketplace bundles is what produces the 70x gap on paper.
Lower the barrier to writing extensions and any ecosystem will balloon just as fast. Claude Code recently opened up its internal script extension system, letting developers write small custom scripts to modify the tool itself. Within 39 days, the community had already spun up 7,848 scripts, nearly half of them with zero stars. Comparing raw plugin totals is meaningless now; what matters is what problems the most popular plugins actually solve.
Breaking Down Account Walls and Cutting Daily Chat Costs
Many developers pay for two or three AI subscriptions at once, yet the apps themselves stay walled off from one another. You cannot spend ChatGPT quota inside Claude Code, you cannot log into a Claude account inside Codex, and Cursor only recognizes its own subscription. All of those tools keep login flows and model selectors locked down, off-limits to third-party plugins.
In DSH, the built-in model connector is just another plugin, which means third-party code can inject new providers directly into the default model menu. Install the cross-platform subscription plugin, and the settings page lets you sign into your own ChatGPT, Claude, Grok, or GitHub Copilot accounts (for Claude, it can even reuse the existing login session on your machine). Once connected, the monthly subscription models from all of those providers show up right inside DSH’s default dropdown menu, with remaining quota for each service displayed live in the bottom status bar.
That convenience, however, relies on identity spoofing. Anthropic’s consumer subscription endpoint only accepts traffic from its own Claude Code client, so to slip past the gate, this plugin forges the latest Claude Code version number in every outgoing request header and injects a fake identity line at the very top of the system prompt claiming to be Anthropic’s official CLI. That directly violates Anthropic’s official legal and compliance terms, which explicitly forbid third-party software from routing requests through personal subscription credentials. In the plugin’s issue tracker, users have already reported that unfamiliar tool names or prompt signatures leaked by other DSH plugins tipped off Anthropic’s servers, triggering immediate blocks that refused to deduct from their monthly plan and demanded pay-as-you-go API credits instead, with heavy usage carrying a real risk of account bans.
To save money on routine tasks, many developers switch to cheap text-only models (such as DeepSeek’s main chat model, which lacks native vision support), only to hit a wall the moment they need to debug a screenshot of an error. In other tools, even if you bolt on an external OCR helper, you still have to save the screenshot as a local file first and pass the file path to the AI. In DSH, installing ModLens, a vision plugin with more than 4,000 stars, adds a vision-badged variant of those text-only models right into the model selector. Pick that option, paste a screenshot straight into the input box, and the chat window renders the thumbnail normally while the plugin extracts the text and layout structure in the background before feeding it to the model.
Multi-day conversations also get expensive fast because every new message resends the entire accumulated chat history back to the model. A long-conversation folding plugin lets the AI collapse finished topics into a few lines of summary, carrying only the summary forward into subsequent turns while letting you expand the full details whenever you need them, cutting out huge amounts of repetitive token spend. In Claude Code and Codex, plugins cannot rewrite the message history sent to the model, so saving that money requires routing traffic through an external proxy outside the app; in DSH, installing the folding plugin handles it natively inside the client.
Editing Remote Code Directly and Rebuilding the UI and Remote Controls on Demand
When your code lives in the cloud or on a remote office machine rather than the laptop in front of you, Claude Code and Codex struggle to work with it directly. Even with Claude Code’s internal scripts, both tools wire their built-in file editing and command execution strictly to the local disk. Reaching a remote machine requires bolting on extra third-party tools that the AI is less practiced at calling and more likely to fumble. DSH instead lets the AI keep calling its familiar built-in file and shell tools while a plugin quietly redirects the actual execution to a remote target in the background. Install the cloud storage plugin Mirage, for instance, and the AI reads and writes cloud files as if they were sitting on a local drive. Install the remote server plugin Honeybee, and your chat history stays on your local laptop while code edits and terminal commands are automatically dispatched to the remote machine.
Beyond cutting costs and reaching remote machines, the look and feel of the application itself can be rebuilt around the task at hand. Although both Claude Code and Codex now ship official desktop apps, and Claude Code’s internal scripts can render buttons, inputs, and simple vector graphics in the sidebar using preset UI components, those scripts are confined to assembling about a dozen preset widgets inside a fixed panel, with total image assets capped at 2 MB and no way to strip away the app’s outer shell; Codex plugins cannot alter UI layout at all. Cursor lets extensions open freeform webviews in the sidebar, but the plugin and the AI live in separate rooms: the panel cannot read the live chat stream next to it, and clicking a button or image inside the panel cannot directly instruct the AI what to edit next.
DSH’s web and desktop builds share the same plugin architecture, allowing plugins to open full graphical workspaces right beside the chat pane and keep them synced with the live conversation state. The multimodal design canvas plugin, for example, lays out web prototypes, slide decks, and video timelines next to the chat. You can drag elements around by hand or click a specific heading or image on the canvas to tell the AI to revise just that single component, then export the result to PDF or presentation slides. Some users have even repurposed DSH for non-coding work entirely: the Shutong Novel Studio plugin strips out the coding layout altogether and replaces it with a bookshelf, outline manager, and side-by-side revision desk, turning DSH into a local long-form fiction editor.
Going the opposite direction, developers who find graphical windows cluttered can install a minimalist terminal plugin that swaps the entire interface back into a keyboard-only dark terminal screen reminiscent of Claude Code, without touching a line of official source code. You can also control your setup from a phone while away from your desk: a web kanban plugin organizes background scheduled tasks into multi-column cards and lets you scan a QR code on your phone to check progress and send instructions on the go, while a messaging bridge plugin hooks DSH directly into 9 chat platforms including Feishu, WeChat, DingTalk, and Slack.
The Flip Side of Teardown Freedom: The Hardest Bill for an Agent OS to Pay
Total teardown freedom has a dark side, and the bill came due on the very day DeepSeek published its milestone numbers. In a reflection post from a power user, the author noted that nearly every installed plugin quietly appends its own block of usage rules to the instructions sent to the AI. Once they passed 15 active plugins, those competing rule blocks crowded the prompt and contradicted one another, sending the AI into endless loops until the user had to delete plugins just to make the tool usable again.
An upgrade failure report on the desktop app posted the same day was even more rattling. After running an in-app update on the Windows desktop beta, a user found that not only did an incompatible third-party plugin prevent the app from launching, the updater had also silently overwritten their configuration file, wiping 34 configured plugins down to just 2 built-in defaults and forcing 5 hours of manual recovery.
Claude Code, Cursor, and Codex are like move-in-ready apartments where the load-bearing walls are sealed shut: low maintenance, but you cannot change the floor plan. DSH is an exposed-conduit shell that lets you knock down even the load-bearing walls and rebuild them yourself. DeepSeek’s push to launch an official plugin marketplace and shore up compatibility and security shows the team has already run headfirst into the real bottleneck: calling yourself the operating system of the agent era is easy, but keeping a wide-open chassis from losing user configs on upgrade or letting plugins poison the prompt is hard, to say nothing of deciding whether an official store can host popular plugins that spoof rival clients to piggyback on competing monthly subscriptions. In our own day-to-day use of DSH, we never pack it with 20 flashy add-ons; installing just one or two plugins that patch specific blind spots in our main coding tools turns out to be the most practical way to use it.