DeepSeek Creator Mode 101: Building Harness Plugins by Chat

DeepSeek Creator Mode is a setting in DeepSeek Harness, DeepSeek’s open-source agent app, that builds a working plugin from a plain-language request. You describe the tool you want, and the agent writes the plugin files, installs them and tests them in the same chat.

DeepSeek’s Harness team showed it off on November 2, 2026, according to TechNode’s report. The demos included a paper reader, a poster designer and a video editor. This piece covers what Harness is, how Creator Mode works, what runs on your computer and what goes to a model provider, and what agencies, nonprofits and campus IT teams should decide before anyone tries it.

The short version: Harness is MIT-licensed and still a preview, and DeepSeek warns it has not had a security audit. Creator Mode turns on a tool that installs plugins, and installed plugin code runs outside the Harness sandbox with your user account’s rights. The app runs locally and can use other model providers. If you use DeepSeek’s own models, DeepSeek’s privacy policy says it stores personal data in the People’s Republic of China. Treat it as a test-machine tool for now.

What DeepSeek Harness is

DeepSeek’s Harness page describes an open-source agent harness for documents, spreadsheets, code and scheduled tasks. It is built on an "everything is a plugin" design, using the open-source Cordis framework. Core features such as the terminal, the agent loop and subagents ship as plugins too.

You can run it three ways, per the same page:

• A desktop app for macOS on Apple silicon or 64-bit Windows.
• A web interface started with npx @deepseek-ai/dsh web, which needs Node.js.
• From source, by cloning the deepseek-harness repository on GitHub.

The repository README says the web interface serves http://127.0.0.1:3080 by default, so it runs on your own machine. The README also calls the project a "developer preview" and warns, in capitals, that there will be compatibility-breaking changes. The latest pre-release on the releases page is v0.2.2-rc.1, dated November 4.

If Harness sounds familiar, it is the same harness DeepSeek used for benchmark runs when it updated V4 Flash, which we covered in DeepSeek V4 Flash now speaks Codex. For the model family itself, see our DeepSeek V4 Pro explainer.

How DeepSeek Creator Mode builds a plugin

DeepSeek’s Harness page walks through one example. A user asks for a Pomodoro timer plugin with Creator mode selected. The agent loads a plugin-development skill, writes a package.json and a client.js file, installs the bundle with its plugin manager tool and checks that it works. The page shows the run taking a little over five minutes.

The plugin manager is the key piece. Its README on GitHub says the plugin_manager tool is enabled in Creator mode, while other presets keep it off by default. That tool can list, install and change plugins for the whole profile. It can install from the npm registry, a local path, a Git repository or a tarball.

There is a check before each install. The same README says every tool action needs either the full-access permission mode or your approval for that call. Dependency build scripts need a separate approval in the conversation. A failed or canceled install restores the previous plugin list.

Try the DeepSeek Creator Mode demo prompts

DeepSeek published a collection of seven demo prompts from its Creator Mode video. The steps on that page are short: copy a prompt, paste it into Harness, select Creator Mode and run it.

The seven DeepSeek Creator Mode demos are a 3D gallery of landmarks, a 3D Transformer explainer, a pixel-style theme for the app, a poster design workbench, a video editing workbench, a paper reader and a conversation statistics panel. The page notes that some plugins were refined over several rounds of conversation, so one prompt may not match the video exactly.

To try it, you need a model. The Harness quickstart guide says to open Settings, then Models, and enter a DeepSeek API key. The demo on DeepSeek’s page shows a DeepSeek V4.1 Flash model selected.

What runs locally and what goes to a model provider

The table sorts the pieces by where they run, based on DeepSeek’s docs and policies.

Harness can run with other models. Its providers guide lists built-in providers such as Anthropic, OpenAI, Moonshot AI (Kimi) and Z.ai (GLM). It also accepts custom endpoints that speak the OpenAI Chat Completions, OpenAI Responses or Anthropic Messages formats. DeepSeek does not say whether DeepSeek Creator Mode works as well with other models.

Privacy and data location

The DeepSeek Harness privacy policy, last updated September 20, 2026, names Hangzhou DeepSeek Artificial Intelligence Co., Ltd. as the operator. It says DeepSeek collects, processes and stores personal data in the People’s Republic of China.

For DeepSeek’s own model service, the policy lists session logs as collected data. That includes your inputs, uploaded files, model outputs and tool call information. The policy says data may be used to train DeepSeek’s models, and it describes an opt-out.

Custom models are handled differently. When you point Harness at your own provider, the policy says Harness "does not store or process your Input and Output on the server side." The provider’s own policy then applies. The policy also says third-party plugins and MCP servers collect and process data directly, and DeepSeek disclaims responsibility for them.

Pricing and license

The code is MIT-licensed, per the footer of DeepSeek’s Harness page and the repository. DeepSeek does not list a price for Harness itself.

Model use is billed separately. The Harness terms of use say DeepSeek’s default models need a prepaid balance on the DeepSeek Open Platform. If you use another provider, you pay that provider. Check DeepSeek’s API pricing page for current token rates, as of November 2026.

DeepSeek Creator Mode plugin risks to weigh

DeepSeek is direct about the risk. The project’s safety notice calls Harness "experimental developer-preview software" with no security audit. It says the app can run model-written code, load third-party plugins and reach the files, network and credentials it is given.

The sandbox has limits of its own. DeepSeek’s sandbox reference says it governs file effects of spawned processes. Network access is outside its scope.

Sharing adds supply-chain risk. Plugins can come from npm, Git or a community GitHub topic, dsh-plugin. The plugin market guide says Harness ships with no index source and checks an index’s file format, not its claims. It tells users to add only sources they trust. The November 4 pre-release renamed this feature "custom plugin index," so menu labels may differ from the docs.

We have seen this pattern with other agent tools. Our post on Plugin4Shell showed a pinned plugin that was not really pinned. The npm supply chain post covers the same trust problem one layer down.

How other platforms handle plugins

OpenAI’s help article on plugins says ChatGPT plugins package reusable instructions, connected tools and other capabilities. Its Plugin Extensions docs let developers place plugins in the ChatGPT sidebar, composer and file viewers. We covered the move in Custom GPT retirement. Anthropic’s plugin admin article lets owners set each Claude plugin to installed by default, available, not available or required. See our Claude for Nonprofits plugin post and what AI skills are. Google’s moves are in Antigravity plugins and Gemini Skills replace Gems.

What agencies, nonprofits and campus IT should decide

Harness and DeepSeek Creator Mode are worth a look for developers who build internal tools. They are not ready for staff laptops. A short checklist:

1. Decide where it may run. DeepSeek’s safety notice recommends a disposable virtual machine, container or dedicated environment, with backups and least privilege. Follow that.
2. Decide which model it may use. Using DeepSeek’s models sends prompts and files to a service that stores personal data in China. If your data rules or client contracts forbid that, set a different provider before the first prompt.
3. Keep client data out. Test with sample files only.
4. Review every install. Read the approval prompt. Use the workspace-write preset, which asks before acting, not full access.
5. Treat shared plugins as code. Read the source of any plugin from npm, Git or an index before you install it. Use https index sources only, as DeepSeek’s guide advises.
6. Expect breaking changes. Pin a version and recheck after each release.

Frequently Asked Questions