Table of Contents
DeepSeek has introduced Creator Mode for DeepSeek Harness, letting users build plugins inside its open-source agent app through conversation. The feature turns requests for custom tools, workspaces, and visualizations into extensions users can run within the application.
The launch demonstrations span a design workspace, a sandbox game, and a 3D language-model visualizer. DeepSeek has also published prompts behind seven demo plugins that users can recreate and adapt.
Creator Mode extends the app's plugin-based design to the process of customizing the app itself. Users can ask the agent to add functionality or change the interface, rather than write an extension manually or wait for an existing add-on.
How Creator Mode Builds Plugins
DeepSeek Harness is the software layer around an AI agent that connects model responses to tools, files, and an interactive workspace. It runs as a desktop application or a web interface and uses Cordis, a framework built around an "everything is a plugin" architecture. Plugins can extend the agent's tools, skills, and interface.
A Pomodoro timer demonstration on the Harness product page shows what that means in practice. After a user requests a timer, the agent loads a plugin-development skill, inspects available interface providers, and identifies a floating overlay as the place to display it.
The agent then writes the plugin's package and client files, installs the bundle through the plugin manager, edits the implementation, and verifies the live plugin state. The resulting timer appears as a floating window with focus and break controls. The interaction combines code generation, installation, and checking the running extension within one conversation.
Underneath that workflow, Harness plugins register capabilities through a context object supplied by the framework. The plugin-development tutorial starts with a TypeScript module exporting an apply function, which the framework calls when loading the plugin. Dependencies can be declared so that required services are available before the plugin starts.
The context also manages a plugin's lifecycle and cleans up registered event listeners, tools, and timers when the plugin unloads. Developers can supply explicit cleanup functions for resources such as network connections. Creator Mode gives users a conversational route into this extension system, while the underlying plugin code remains part of the implementation.
Seven Prompts Cover Design, Research, and Visualization
DeepSeek's English and Chinese demo prompt collection shows the range of changes users can request. To try an example, users copy its prompt into Harness and select Creator Mode.
The seven examples cover both standalone workspaces and modifications to the host app:
- A 3D landmark gallery adds a left-sidebar entry with models of the Eiffel Tower, the Great Wall of China, the Forbidden City, and another landmark the agent chooses.
- A Transformer visualization is an interactive 3D demonstration game meant to explain the computation inside a Transformer, the neural-network architecture used in many language models.
- An interface-theme switcher changes the app's appearance across five styles, including a pixelated look.
- A poster workbench opens a design tool from the conversation that uses the Harness website's shader background and English font and has editable titles, text, and shader colors.
- A video-editing workbench is a dark-themed editor with a large preview and timeline; its prompt requests support for multiple video uploads and export and natural-language editing through the conversation.
- A paper reader sits in the right sidebar, and its prompt asks it to preserve a paper's layout and provide highlighting, notes, and questions about selected text.
- A conversation-statistics dashboard is requested to use actual conversation data for token totals, conversation counts, estimated costs, budgets, usage timelines, and CPU activity.
DeepSeek notes that some examples were refined across several rounds of conversation, so a single prompt may not reproduce the exact result shown in the video. Users can continue the conversation to adjust the generated plugin.
Running the Open-Source Harness
DeepSeek Harness is in public preview worldwide, with its source code available under the MIT license. Alongside Creator Mode, it supports document work, data analysis, coding, research, and background tasks. Its official plugin interface includes terminal, agent-loop, and subagent components, plus experimental agent teams, approval review, scheduling, and voice input.
For the web version, users install Node.js and run npx @deepseek-ai/dsh web. The command starts a local web interface, using port 3080 by default.
The web setup guide then directs users to enter a DeepSeek API key under Settings → Models and select a workspace directory before starting a session. From that workspace, the agent can read and edit files, execute commands, and run tasks, with approval requests governed by the active permission policy.
