DeepSeek Harness hits 1.5 million daily users and prepares an extension store for AI agents

3–4 minutes

DeepSeek’s open-source platform, used to create and run artificial intelligence agents, reached 1.54 million active daily users in less than three months. The Chinese company announced this figure on Sunday, October 8, and it was reported by the South China Morning Post. The data refers to users who accessed DeepSeek’s official processing service on November 4. At the end of September, daily users numbered around 271,000, according to The Business Times.

Fueled by this rapid growth, the startup plans to launch an official store for extensions that add new features to agents. The community has already developed around 35,000 extensions, with approximately 60% of active users creating or installing at least one, according to DeepSeek. The store’s launch date has not yet been announced.

Harness positions DeepSeek in a competition that goes beyond chatbots. Like Anthropic’s Claude Code and OpenAI’s Codex, the platform enables AI to open files, modify code, and run tests. With extensions, the company aims to help developers adapt agents to various work routines, from reading scientific articles to producing content.

How Harness manages tasks between responses

Models from the DeepSeek V4 family interpret user requests and can determine which action to execute. Harness organizes the execution, provides the agent with access to enabled tools, records what happened, and returns the result to the model, which decides whether to continue or terminate the process.

The same logic allows an AI agent to move from one file to a terminal or a search, within the access permissions it has received.

In DeepSeek’s tool, developers can replace or add components such as the model used, how session information is stored, and the programs the agent can invoke. One project cited by the company combines reading, translation, and annotation of scientific articles in a single environment.

Another example shows what information the agent received at each stage of a task, allowing the user to trace the origin of a decision.

In late October, the team released a prototype demonstrating how far the company plans to take this idea in programming. The system monitors a GitHub repository and, when it receives a request from an authorized person, can work in its own session and return a code change proposal for review.

DeepSeek describes the project as a research prototype and notes that some of the discussed functionalities are still under development.

Why the extension store is strategic for DeepSeek

Anthropic already offers an extension catalog for Claude Code, with integrations to services like Slack and Figma. OpenAI also provides extensions for Codex. The value of these tools becomes clear when they are integrated into existing team workflows. For example, Oracle now allows developers to use Claude Code and Codex to create applications with agents within its enterprise management system.

DeepSeek can compete in this market in two ways. With Harness, the company aims to encourage users to open their own applications for work. Through the API, it can sell processing power even when developers prefer another interface.

The company provides instructions for connecting its models to Claude Code and Codex. This allows a team to retain the tool it already uses while replacing the model that processes certain tasks.

Usage volume helps explain the commercial interest. On November 4, Harness recorded around 92 trillion tokens per day on the official API, according to DeepSeek. Tokens are the units into which the model divides the content it receives and produces.

A single task can pass through the model multiple times and include text from files, instructions, and test results. The volume indicates the amount of content circulating through the system. To calculate revenue, it would be necessary to know how much of this usage was billed and at what rates, which vary by model, time, and type of processing.

DeepSeek says it plans to improve the compatibility and security of extensions, as well as offer processing credits and technical support to creators. Harness documentation still classifies it as a preview for developers and warns that updates may affect existing extensions. For a team that deploys an agent to work on files and internal systems, controlling access and maintaining integrations is as important as the quality of the code the model produces.