DeepSeek's product team has published fresh numbers for DeepSeek Harness, its Agent tool. As of Nov. 4, and counting only users who call the official API, Harness had 1.54 million daily active users and consumed about 92 trillion tokens per day. The number of third-party plugins has grown to roughly 35,000, and about 60% of active users have built or installed at least one of them.
The product's timeline is short. v0.1 went live on Aug. 13 and the desktop version was formally released on Sept. 29, which puts this disclosure less than three months after launch.
What plugin authors are building
The official post highlights four community plugins, and they point in very different directions:
- A workbench for social media creators (dsh-media-desk), which puts note-taking for ideas, script writing, cover design and storyboarding in one interface;
- A design plugin (dsh-guizang-design), which merges design and copy-polishing Skills into one plugin. Articles can be turned into web pages, image-and-text cards or covers, and a "make a similar one" feature lets users reuse a result;
- A context-insight plugin (dsh-context), which shows users what information the Agent drew on while working and how that information shifts over the course of a conversation;
- A paper-reading plugin (agentero-dsh), which keeps reading, highlighting, translating and note-taking in one place.
Guizang, the author of the design plugin, said the plugin system "can offer much more than Skills" and lets works, columns and a personal tone accumulate over time. Judging by these examples, people are using Harness for far more than writing code, with content creation and academic reading taking up a good share.
How big is 92 trillion tokens a day?
A rough calculation with the two figures: 92 trillion divided by 1.54 million works out to about 60 million tokens per daily active user per day. That average is very high, which suggests that Harness mostly runs long-context, multi-turn Agent tasks; single question-and-answer exchanges could not account for that volume. The average is also likely pulled up by a small number of heavy users and automated workflows. DeepSeek has not published the distribution, so the per-user figure should be read only as an order of magnitude.
The counting basis deserves attention as well. DAU includes only users on the official API, so people who connect third-party APIs are not counted. The post does not say separately whether the token consumption figure uses the same basis.
Next up: a plugin marketplace and author incentives
DeepSeek listed several plans: improving the stability and compatibility of the plugin API, strengthening plugin security mechanisms, and building an official plugin marketplace where authors can upload and users can search. A support program for plugin authors is in preparation, with token incentives, technical exchange and work recommendations. Startups and enterprises will also be able to connect their own products and services through plugins. Current sign-ups are only an early expression of interest, and the specific rules will be announced when the program formally launches.
DeepSeek positions Harness as "the operating system of the Agent era". It plans to build only the stable foundations and the general features most people need, leaving specialist scenarios and industry needs to plugins. Under the "everything is a plugin" design, even built-in features can be replaced by third-party plugins.
Plugin security will be an unavoidable hurdle for this approach. Earlier this year we reported on the supply-chain trust crisis around the open-source Agent framework OpenClaw, and on a fake Claude Code plugin that slipped into Cursor and was installed by up to about 5,000 people. The more plugins there are, the more pressure falls on review and permission isolation. DeepSeek only said it would "improve plugin security mechanisms" and has not yet disclosed details of its review process, permission tiers or signing mechanism.
Sources: CocoLoop, official article from the DeepSeek product team; we checked the DAU counting basis, average daily token consumption, plugin count and product launch milestones.