DeepSeek Harness daily users hit 1.54 million in under three months

Felix Laurent 8 Nov 2026 4 min read

AI agents

DeepSeek Harness daily users reached 1.54 million on November 4, the company said. The open-source agent app from the Chinese AI lab launched less than three months earlier, on August 13. At the end of September, it had about 271,000 daily users.

DeepSeek first shared the numbers in Chinese on November 5 and published an English post on X on November 8. Bloomberg reported the same day that the figures show DeepSeek challenging Anthropic and OpenAI with a growing set of open-source agent tools.

How fast DeepSeek Harness daily users grew

Most of the growth came in October. The figures below come from the Harness team, as reported by PANews and Bloomberg.

Date	Daily active users	What happened
August 13	–	Developer preview v0.1 goes live
September 29	–	Desktop app released
September 30	About 271,000	7-day average
October 7	About 429,000	7-day average
November 4	1.54 million	Single day

That is more than a million new daily users in one month. The same users consume about 92 trillion tokens per day, according to DeepSeek. The jump started right after the desktop app came out, which made Harness usable without a terminal.

The headline number needs some context, though.

What the 1.54 million figure counts

DeepSeek counts only people who run Harness through its official API. Harness can also work with other models, so users on other providers are left out. The company has not shared how many of them there are.

The two earlier figures are 7-day averages, while the November number is for a single day. The data is self-reported, and no outside firm has checked it. Direct comparisons with rivals are also hard. OpenAI and Anthropic usually report weekly users for Codex and Claude Code, and they rarely give daily counts.

Other signals point the same way. The GitHub repository has about 276,000 stars and 34,000 forks. That puts it among the most starred AI projects of the year. Much of that interest comes from how the app is built.

Why developers pick DeepSeek Harness

The app uses an "everything is a plugin" design. The model adapter, the tools and even the agent loop can be swapped for a third-party plugin. It is released under the MIT license and is still labeled a developer preview.

The plugin count has grown to about 35,000, and about 60% of active users have built or installed at least one. DeepSeek highlighted a few of them:

dsh-media-desk: a creator workbench for scripts, covers and storyboards;
dsh-guizang-design: turns an article into a web page, an image card or a cover;
dsh-context: shows what context an agent sees and how it changes;
agentero-dsh: reading, translating and annotating research papers with an agent.

Some plugins turn Harness into a hub for rival tools. InfoQ, in a piece republished by 36Kr, describes dsh-aio-agent, which lets users run Claude Code, Codex, Cursor and Kimi Code from one interface. The team is also testing more ambitious ideas, such as the proactive super-agent it published on November 2.

Freedom has a price, and users have started to report it.

What users complain about

The same InfoQ piece collects user complaints. One developer installed more than 15 plugins and ran into infinite loops, context overflow and rising token bills. Each plugin adds descriptions and rules that the model must read before it works.

Another user said a small project cost more than 20 times the tokens it took in OpenCode. The task took 30 to 40 minutes against about 30 seconds. That is one user's experience, but the token concern is common in the feedback.

DeepSeek has shipped fixes at a fast pace. Updates in October and early November added plugin start and stop, dependency sync, hot reload and automatic compression of long conversations. A developer statistics plugin now shows how much context, cache and tool time a task uses.

Those fixes set up the next step of DeepSeek's plan.

What DeepSeek plans next for Harness

DeepSeek wants Harness to become "the operating system of the agent era". The company plans an official plugin marketplace and a support program for plugin authors. It will offer token grants, technical help and promotion, but the rules and dates are not set yet. It also invites startups to bring their products into Harness as plugins.

The push comes with plenty of money behind it. In October, DeepSeek raised about $15 billion from investors including CATL and Tencent. That came two days after the US ordered federal agencies to remove DeepSeek products. The ban covers only federal civilian systems, so developers elsewhere can still use Harness.

Is DeepSeek Harness a real threat to Claude Code and Codex

It is already a serious rival in reach, if not yet in polish. Getting 1.54 million people to use an agent app daily within three months is fast for any developer tool. It is especially fast for one still in preview. The open plugin model is also hard for Anthropic and OpenAI to match. Users can rebuild the whole agent around their own work.

The weak spots are just as clear. The numbers are self-reported and cover only DeepSeek's own API. Users also complain about token costs and instability. If the marketplace makes good plugins easy to find and keeps bad ones out, Harness could hold its users. If not, its growth may stall where the plugin sprawl starts.