DeepSeek Harness gets Creator Mode, which builds plugins from a single sentence

What's going on

On November 2, 2026, Hangzhou-based Chinese AI company DeepSeek unveiled a new feature for its agent software DeepSeek Harness: Creator Mode. You describe the tool you need in plain language. The AI agent then codes it as a plugin, an add-on module that hooks directly into the Harness app. It installs the plugin itself and checks that it works. DeepSeek is promoting the feature on social media with the slogan "Your Harness. You Define It." Demos so far include a design workbench, a sandbox game and a tool that visualizes large language models in 3D.

Harness itself is an open-source agent framework. That's the scaffolding that turns a language model into an assistant that can act on its own. It's released under the permissive MIT license, and DeepSeek says it's currently in a public beta worldwide. You can run it as a desktop app for Macs with Apple chips and 64-bit Windows, or as a local web interface in your browser.

Seven examples, from 3D landmarks to video editing

For the launch, DeepSeek published seven sets of example prompts in Chinese and English. They're meant to show that Creator Mode can do more than build small helper tools:

The idea is that nobody has to learn Harness plugin development first. You describe the function, interface and controls, and Harness tries to turn that into a plugin. But in the prompt collection, DeepSeek itself admits that some demo plugins were refined over several rounds of conversation. So a single prompt won't necessarily reproduce what you see in the videos. Creator Mode lowers the barrier to entry, but it won't reliably produce complex apps at the push of a button.

How it works: the agent extends itself

The foundation is an architecture DeepSeek built earlier. Harness is based on Cordis, a plugin system, and follows the principle "Everything is a plugin": models, tools, skills and even parts of the interface are swappable building blocks rather than hard-wired features.

On its product page, DeepSeek walks through the process using a Pomodoro timer, a timer for focused work sessions:

1. The agent loads the Cordis skill for plugin development.
2. It checks the runtime environment and the available APIs.
3. It writes the plugin's configuration and code files.
4. It calls the plugin manager's install tool.
5. It checks whether the plugin actually runs.

The result is a floating timer with "Focus" and "Short break" buttons. It shows up in the list of installed plugins and is reported as built, installed and verified. According to the demo, the whole thing took about 5 minutes and 24 seconds.

That's what sets Creator Mode apart from ordinary code generation in a chat. Instead of a code snippet or a single web page, the agent expands its own working environment. According to Unite.AI, DeepSeek laid the groundwork step by step. Version 0.1.6-alpha.2, released on September 17, switched to persistent plugins that are installed through a plugin manager and can be toggled on and off live. The 0.2.1-alpha.1 pre-release from October 3 added a "Let Agent create a plugin" entry to plugin management that jumps straight into Creator Mode. An experimental component for custom plugin marketplaces also adds a "Discover" tab. However, DeepSeek doesn't ship a source of its own for it, and only the file format gets checked. Titles, author details and the origin of the listed plugins are not verified.

Background: DeepSeek takes aim at Claude Code

According to Bloomberg (cited by Quartz and Shiftdelete), DeepSeek set up the Harness team in early 2026 to keep pace with coding agents from US companies. The main rival is Anthropic's Claude Code. So far, Harness is aimed mainly at developers rather than everyday users. Besides coding, it's meant to analyze data, organize files and do research.

Creator Mode is one of a whole series of updates. In October, according to Quartz, DeepSeek optimized its V4.1-Flash model specifically for Harness and recommends Harness 0.2.1-rc.1 or later for it. A few days before Creator Mode, the company showed off a prototype called "Super-agent" and an improved coding assistant that's supposed to spot problems early and need little human supervision. DeepSeek is in good shape financially: according to Bloomberg, it raised around 100 billion yuan (about $15 billion) in its latest funding round and is preparing an IPO. According to Shiftdelete, the Harness project has picked up more than 269,000 stars on GitHub since August.

What you need to try it

Harness is an app, not a model. So you don't need a powerful graphics card, because the language models are usually connected through an API. There are three ways to run Harness:

• Desktop app for Macs with Apple chips and 64-bit Windows.
• Web interface: Install Node.js, then run npx @deepseek-ai/dsh web. The interface opens locally at http://127.0.0.1:3080 in your default browser.
• From source: Clone the repository, then run pnpm install, pnpm run build and pnpm dsh web.

Read the security notes. DeepSeek explicitly calls Harness experimental software that has not undergone a security audit. It can run AI-generated code and commands, load third-party plugins, and access the network, processes, credentials and any files you give it access to. Faulty output, malicious input or untrusted plugins could delete files or leak data and passwords. Sandboxing (isolating software in a protected environment), approval prompts and permission controls reduce the risk, but DeepSeek says they don't guarantee protection. Its advice: grant as few permissions as possible, ideally run it in a disposable VM or a container, keep backups, and review plugins before running them. DeepSeek also warns of breaking changes ahead, so plugins you build today may need reworking later.

Our take

Creator Mode shows where AI agents are heading: from using tools to building them. Until now, development teams decided what an agent could do. Here, its feature set grows with what users ask for. For DeepSeek, there's a strategic bonus. If users and developers build their own plugins, an ecosystem emerges that the in-house team could never cover alone. The fact that it's all open source under the MIT license sets Harness apart from closed rivals like Claude Code.

What's still unclear is whether auto-generated plugins can become stable working tools. With complex plugins, correct code, compatibility, data access and security are hard problems. DeepSeek's own admission that the demos were refined over several rounds tempers the "one sentence, one plugin" promise. The plugin marketplace is especially risky because nobody verifies where plugins come from. A malicious coding extension that recently slipped in an SSH backdoor showed how quickly tampered extensions can do damage. If you're a tinkerer with an isolated test environment, Creator Mode is a fun toy. For production use with real data, it's still too early.