DeepSeek's open coding agent reports sixfold jump in users, challenging Claude Code and Codex

DeepSeek, the Chinese artificial intelligence company, said this week that daily use of its open-source coding agent framework has grown roughly sixfold since late September, the first time the company has disclosed usage figures for the product since it launched in August. The tool, called DeepSeek Harness, is built to compete directly with Anthropic's Claude Code and OpenAI's Codex, the dominant paid tools that let AI models write, test and fix software with minimal human supervision.

The disclosure, posted by DeepSeek's Harness team on the evening of Nov. 5 and reported the following day by the Shanghai-based financial news outlet Yicai Global and separately by the Chinese finance site Gelonghui, landed in Western technology coverage this past weekend as Bloomberg News picked up the figures. It arrives as US and Chinese AI companies compete openly for the allegiance of software developers, a group increasingly asked to delegate entire coding tasks to autonomous AI agents rather than writing code line by line.

A fast climb in five weeks

According to the figures DeepSeek released, daily active users of Harness on the company's official application programming interface reached about 1.54 million as of Nov. 4. That compares with roughly 429,000 for the week ending Oct. 7 and about 271,000 for the week ending Sept. 30, both outlets reported. Users of the tool were consuming roughly 92 trillion tokens — the unit AI companies use to measure text processed by a model — per day by early November, according to the same disclosure.

DeepSeek also said roughly 35,000 third-party plugins have been built for Harness since launch, and that about 60 percent of active users have installed or created at least one, Yicai Global reported. The company said it is preparing a support program for plugin developers and an official plugin marketplace, though it gave no date or terms.

The figures are self-reported by DeepSeek and track usage only through the company's own official API; they are not independently audited. Yicai Global's own report noted that the number should not be read as the total independent user base of the separate Harness desktop client, since DeepSeek's broader consumer chat application has a far larger following. No outside analytics firm has yet published a comparable, independently measured figure for Harness.

Built to challenge two American products

DeepSeek Harness is what the industry calls an "agent harness": software scaffolding that wraps a large language model with tools, memory and permissions so it can carry out multi-step coding work — fixing bugs, writing tests, refactoring code — over extended stretches without constant prompting. Anthropic's Claude Code, which launched in February 2025, and OpenAI's Codex, introduced months later, established that category and have built paid subscriber bases of professional developers around it.

DeepSeek began assembling a team earlier this year explicitly to build a domestic rival to those products. Harness shipped as a developer preview on Aug. 13 under an open-source MIT license, with a desktop client following on Sept. 29. DeepSeek's own API documentation still describes the product as being "in developer preview," warning that "compatibility-breaking changes" should be expected. The project's code is hosted on GitHub, where it has drawn more than 270,000 stars, a rough measure of developer interest, though not of actual usage.

The push accelerated in late October. DeepSeek reorganized its product division on Oct. 16 to bring consumer-facing AI tools, including Harness, under a single leader, according to Chinese media reports. On Oct. 31, the Harness team published a technical post describing a "super-agent" architecture meant to let an agent hold context across sessions and juggle multiple tasks at once, rather than resetting with each new request. Days later, the team released an early, open-source prototype called Super-Agent that monitors developers' group chats and proposes fixes to bugs on its own, rather than waiting to be asked — a different approach from the request-and-respond model used by Claude Code and Codex, Digital Trends reported on Nov. 1.

Developers, and two incumbents, take notice

The most immediate audience is software developers deciding which AI coding tool to adopt, particularly outside the United States, where Harness is free to run against DeepSeek's own models and can be pointed at other providers' models as well. Open-source infrastructure teams, students and companies wary of per-seat subscription costs for Claude Code or Codex are the likeliest early adopters of a free, self-hostable alternative.

For Anthropic and OpenAI, the figures are notable less for their scale — 1.54 million daily users is modest next to the tens of millions who use DeepSeek's main chatbot — than for the trajectory. A sixfold rise in five weeks, even from a small base and even if not independently verified, suggests DeepSeek's push into coding agents is gaining real traction rather than stalling after launch, as some early reviewers had predicted. Several technology outlets that tested Harness over the summer described it as rougher around the edges than Claude Code or Codex, requiring more configuration to match their polish.

Reaction so far has been measured rather than alarmed. Coverage of the Super-Agent prototype cautioned that an early, open-source release is not evidence DeepSeek has overtaken its American rivals on reliability or deployment readiness, and analysts who cover the Chinese AI sector have noted that DeepSeek's own disclosures remain the only source for the usage numbers. DeepSeek did not respond to a request for comment routed through its public channels.

What happens next

DeepSeek has not set a date for either the promised plugin marketplace or a stable, non-preview release of Harness, leaving open how quickly the tool might graduate from developer preview to something enterprises are willing to deploy on production codebases. The company's broader ambition, as described in its own technical writing, is to turn Harness into something closer to an operating system for autonomous coding agents, built around a marketplace of interchangeable plugins rather than a single fixed product.

The contest will also play out against a backdrop of US scrutiny of Chinese AI tools in sensitive software environments, a concern that has already shaped how some Western companies evaluate open-weight Chinese models. Whether that caution extends to a coding-agent framework that developers can inspect, fork and run entirely on their own infrastructure — as Harness's MIT license permits — is likely to become part of the argument DeepSeek makes to developers weighing it against Claude Code and Codex in the months ahead.