DeepSeek's Harness Hit 1.54 Million Daily Users in Under Three Months

DeepSeek said on 8 November 2026 that its open-source agent app, DeepSeek Harness, reached 1.54 million daily active users as of 4 November. Version 0.1 went live on 13 August 2026, so the figure landed less than three months after launch. The company announced the number on X, and Bloomberg reported it the same day.

Related reading: OpenAI Moves Frontier Training Off the Live Internet and Onto a Cached Copy of the Web · Anthropic's Claude Duet Works in the Background, and Asks Before It Acts.

The same disclosure put token consumption at about 92 trillion a day and the third-party plugin catalogue at roughly 35,000. Both are DeepSeek's own figures. The user count also measures one route into the product: people connecting through DeepSeek's official API, not everyone running Harness.

That distinction carries more weight than it usually would, because Harness is MIT-licensed. Anyone can self-host it, fork it, or offer it through another provider. The 1.54 million is a floor with an unknown distance to the ceiling, and DeepSeek has not said how far apart the two are.

What an agent harness is for

A harness is the program around a model that gives it somewhere to act. The model writes text. The harness supplies the shell command, the file, the browser tab, the session that remembers the previous step, and the permission check that decides which of those the model may touch. Claude Code and OpenAI's Codex CLI are harnesses in this sense. So is Harness.

DeepSeek shipped it on 13 August 2026 alongside V4-Pro on its API, VentureBeat reported at launch, and the same release came with higher API prices. The New Stack covered the launch and led on the plugin architecture. The app is written in TypeScript and published under the MIT licence, a different posture from the closed runtimes sold by Anthropic and OpenAI.

DeepSeek's product page presents Harness as a general agent client: run tasks, load skills, connect outside services. The launch invited the Claude Code comparison, and the licence is what separates the two. A competitor can read the code, change it and ship the result without asking. The architecture is model-agnostic as well, so a runtime built this way can in principle drive another provider's model when a plugin supports it.

That changes what DeepSeek is selling. If the harness can drive someone else's model, the durable asset is the harness and its plugins, not the model underneath. The model becomes a component the runtime can swap. DeepSeek's models remain the default and its low API prices are part of the pitch, but the architecture does not require them.

Everything is a plugin, including the loop

Harness runs on Cordis, a micro-kernel framework. Models, tools, skills, sessions, sandboxes, storage, the agent loop and the user interface are all plugins. DeepSeek's phrase for the design is that everything is a plugin, and there is no privileged core.

The choice is unusual for a runtime backed by a model lab. Vendors normally keep the scheduler, the interface and the permission layer inside the product and expose extensions at the edge. Here a third party can replace the parts that decide how the agent behaves. DeepSeek's own interface has to compete inside the same system as everyone else's.

The plugin count follows from that openness. About 35,000 third-party plugins exist, and about 60% of active users have built or installed at least one, according to the company. DeepSeek also says the plugins "vary in maturity", which is a polite way of describing a catalogue that is not a quality mark. Thirty-five thousand entries is a supply figure, and nothing in the disclosure says how many work or are maintained.

MLQ.ai's account of the disclosure names three: a creator workbench, a design plugin that turns articles into web pages and image cards, and a literature tool that links reference management, translation and note-taking. None is a core feature of an agent runtime. Each is work a user would otherwise do in a separate application, which is the argument for letting outsiders build them.

The early attention was large. The GitHub repository showed about 95,386 stars and 8,826 forks by 15 August, two days after launch, and about 193,700 stars by 25 August, according to Flowtivity's analysis of the repository. Those figures move quickly, and they measure interest in the project, which is a different thing from use. They do show that the open-source framing found an audience before the user numbers did.

Ninety-two trillion tokens a day

The token figure says more about the workload than the user count does. DeepSeek reports roughly 92 trillion tokens a day across those 1.54 million daily users. Divided straight, that is about 60 million tokens per daily user per day, which is not a chat pattern.

Chat spends tokens when a person types. An agent loop spends them on tool calls, on re-reading context it already holds, on retries after a failed step, and on the next plan. A task that takes forty steps can consume more than forty conversations, and the tokens are spent whether or not the task succeeds. DeepSeek has not broken the 92 trillion down by use case, so how much of it is coding, research or something else is unknown.

Comparisons across providers do not survive contact with the details. Tokenizers differ, and providers count cached or repeated context differently. The 92 trillion is a DeepSeek number under DeepSeek's method, and no rival has published a figure on the same basis. The figure works as a scale indicator; it is not a league table.

The growth curve has its own caveat. Daily users were about 271,000 at the end of September 2026, just after the desktop app shipped on 29 September, according to Bloomberg. That figure was a seven-day average at 30 September, not a single-day count, so the move to 1.54 million is not like for like. It is still more than five times the September number in about five weeks, and the desktop release sits at the start of that run.

The numbers are DeepSeek's, and they count one door

The monthly figure is the one quoted alongside the daily number. FourWeekMBA, reading the same disclosure puts monthly users at 5.48 million and attributes the figure to the company. It is not an independently verified audience measurement, and it does not settle how many of those people pay for anything.

What DeepSeek did not disclose is as telling as what it did. There is no revenue figure, no split between free and paid usage, and no churn or retention rate. A daily-user number without retention says how many people turned up. It does not say how many stayed. For a product that is free to run and MIT-licensed, the gap between 1.54 million users and a business is the whole question.

The token volume does not close it. Traffic through the API is billed at DeepSeek's list prices, but the company has not said what share of that traffic is paid, discounted or covered by credits. Treating 92 trillion tokens as revenue at list price would be an assumption dressed as arithmetic, and the Seeking Alpha report on the disclosure repeats the same set of figures without filling that gap.

There is also a definitional problem with counting users of an open-source app at all. A fork can be renamed, a self-hosted instance can serve a team, and a plugin can wrap the runtime inside another product. None of those users appears in DeepSeek's number, and none is visible to an outsider. The official-API count is best read as a proxy for the company's own distribution, not for the size of the category.

An operating system needs an app store

DeepSeek's stated goal is to make Harness "the operating system of the agent era". The marketplace is the part of that plan with a shape. The company says it will open an official plugin marketplace and a support programme for plugin authors, with token grants, technical discussions and promotion. It is collecting expressions of interest now. Rules and dates will be announced at launch, and no date has been given.

The platform argument explains why the plugin count matters more than the daily-user count in the long run. A model can be copied. An ecosystem of 35,000 plugins, however uneven, is harder to rebuild, because it represents work done by people DeepSeek does not employ. The token grants put a small amount of money behind that work, and promotion costs DeepSeek less than cash.

"The operating system of the agent era." — DeepSeek, Harness announcement, 8 November 2026

The strategy extends a pattern. DeepSeek's open-weight models and low prices made it a force in the model market; Harness carries the same approach up a layer, into the runtime where agents do their work. Our report on DeepSeek V5 and Huawei's Ascend 950 covers the model side of that push. Owning the agent layer would give DeepSeek something a model release does not: a place where users, plugins and workflows accumulate.

It also puts DeepSeek in direct competition with Anthropic and OpenAI where their agent products are strongest. Claude Code and Codex CLI have the advantage of tight integration with their own models and the teams that tune them. Harness answers with openness and price, and with the promise that a plugin can change anything, including the loop. Whether developers prefer to build inside a closed runtime or on top of an open one is unsettled.

The security question sits underneath. A plugin in Harness runs with the agent's permissions, which means it can read what the agent reads and act where the agent acts. With 35,000 extensions from unknown authors, the supply chain is the weak point, and DeepSeek has not described a review process, a signing requirement or a removal policy. Our coverage of OpenAI's GPT-6 cyber-critical launch covers the permissions problem in more detail.

What to watch

The marketplace launch will answer the first question: what DeepSeek requires of a plugin before it lists one. Review, signing, permission scopes and a removal policy are the details that decide whether the catalogue becomes an asset or a liability. The terms will also show whether DeepSeek intends to take a cut, and from whom.

Then there is the metric. A total-user figure, including self-hosted and third-party traffic, would turn 1.54 million from a floor into something closer to a measurement. A retention number would say more than either. DeepSeek has published neither, and it has not said whether it will.

Watch the competitive claims too. Neither DeepSeek nor its rivals have released audited, comparable usage figures for agent runtimes, so any statement that one has overtaken another is unsupported for now. The plugin economy is the number to watch, and nobody outside DeepSeek can currently see all of it.

Token grants will show up as a cost line if the programme scales, and the first serious plugin incident will show up as a security story if the review process is thin. Both are measurable. Neither has happened yet. DeepSeek has not said when the marketplace will open, what it will cost, or how it will police what goes into it.

Cover image: The DeepSeek mobile app home screen. DeepSeek, from the official app.