DeepSeek shares seven Creator Mode examples for building plugins through chat

DeepSeek showcased Creator Mode for DeepSeek Harness on November 2 and published seven example prompts for building plugins. Users describe a tool in conversation, and the agent creates an extension for its workspace. The announcement highlights a design workspace, a sandbox game and a 3D language-model visualizer.

Plugins extend the tools, skills and interface available inside Harness. The product page presents two paths: installing existing plugins and asking the agent to create one through conversation. Users can request an interface suited to their work within the agent environment. Harness remains a developer preview available worldwide.

Seven examples, with a reproduction caveat

The video-editor example requests multiple uploads, a preview, a timeline and export. The paper-reader prompt asks for preserved layout and highlighting, notes and questions on selected text. These are requested features.

DeepSeek says some demo plugins were refined through several conversations. A single prompt may therefore differ from the video. Published requirements still need checking against the resulting plugin.

The product page also illustrates a focus timer. Its displayed sequence moves from writing files to installing a plugin and checking its runtime status. It is a company demonstration, rather than an AI Ping performance test.

Models are configured separately. Official documentation describes connecting a DeepSeek API key, adding another model provider or using a compatible endpoint on a company gateway or self-hosted server. The published Harness source and the access conditions of the chosen model service are separate considerations. The showcase does not announce new model weights.

Distribution has its own mechanism. An experimental plugin-index feature adds a discovery interface, while Harness ships without an index source. User-added indexes supply installable versions and their origins, with confirmation before installation.

Recommended versions account for compatibility with the running Harness. The confirmation identifies the package or Git source. Third parties supply descriptions and author information; checking the index format does not verify those claims.

Plugin changes can affect more than the conversation that requested them. The manager documentation says changes apply to every session using the same profile. Agent management operations require full access or approval for that call. Installed host code runs outside the workspace sandbox, making the scope of an extension broader than its visible interface.

The management interface also separates installation from activation. A completed installation offers a choice to enable the plugin; configurations without immediate reloading require a restart before saved changes take effect. The manager also distinguishes a saved enabled selection from successful loading. A broken plugin retains an error and can be disabled; an installed package alone does not establish successful execution.

DeepSeek consequently describes Harness as experimental software that has not undergone a security audit. It can execute generated code and commands and access the files, network and credentials made available to it. Sandboxing and approval prompts reduce risk without guaranteeing isolation. The company recommends restricted access and a separate test environment.

The public repository carries an MIT license alongside a developer-preview warning about changes that can break compatibility. The product page likewise says core plugins and basic APIs continue to evolve. The newly published examples show ways to extend tools within that changing environment.