DeepSeek Harness Introduces New Plugin Creation Feature

New Features in Plugin Management

On October 3, 2026, DeepSeek unveiled the prerelease of DeepSeek Harness v0.2.1-alpha.1, featuring an exciting addition to its plugin management system. Users can now enable a 'Let Agent create a plugin' option that transitions existing drafts into Creator mode, pending user approval for the plugin creation process.

Updates from Previous Versions

This latest update builds upon earlier features documented in version v0.1.6-alpha.2, released on September 17, 2026. In that release, DeepSeek moved away from the Cordis dynamic execution tools and instead focused on implementing persistent plugins via the Plugin Manager. This included the introduction of a dedicated Plugins page for managing installations, alongside capabilities to enable or disable plugins in real-time and unload them from runtime.

The prerelease in October retains the accessibility of Standard, Creator, and custom modes even with the coding view disabled. By default, built-in configurations—PTC and Minimal—set to Standard mode ensure continuity for existing sessions while adapting to any new settings. Additionally, improvements to plugin bundles address issues such as missing dependencies when a new bundle is activated and stale nodes after bundles are deactivated or removed, though a restart remains necessary to update installed package versions.

Exploring Creator Mode

DeepSeek has positioned Harness as an open-source agent framework, currently available in worldwide public preview. Its architecture, derived from Cordis, incorporates modular plugins, allowing users to either install existing extensions or create their own via chat in Creator mode to enhance tools, capabilities, and user interface.

An illustrative demonstration on the product page showcases a Pomodoro timer as a practical application of Creator mode. Here, the agent activates the Cordis plugin development skill, reviews available providers and runtime slots, and generates necessary package and client files. Following this, the plugin manager’s bundle installation tool is called into action to review the current plugin state. The conclusion of the demo sees the floating timer integrated under Installed plugins, including controls for Focus and Short Break, signifying a successful build, installation, and verification process.

Access and Plugin Functionality

For those interested in exploring the functionalities further, DeepSeek’s project repository is available under an MIT license, with comprehensive Web UI setup documentation accessible via npm or source. Users need to have Node.js installed, and can get started by executing npx @deepseek-ai/dsh web to launch the interface, which defaults to http://127.0.0.1:3080 in their browser. Cloning the repository via the source route involves several commands: pnpm install, pnpm run build, and pnpm dsh web.

Currently, Harness is in a developer preview stage, with potential compatibility-breaking changes anticipated as core plugins and APIs continue to evolve. An experimental Custom plugin market is also documented, which adds a Discover tab and a view-only plugin_market utility within Creator mode. However, disabling this bundle removes both features. Installing discovered versions requires user confirmation before proceeding with the usual compatibility and registry checks.

The validation process within DSH ensures the accuracy of the index file format but does not authenticate the titles, descriptions, icons, authors, or sourcing claims provided by the market.

Safety Considerations

A safety warning accompanies the release, highlighting that Harness is still categorized as experimental software in developer preview, thus has not yet undergone comprehensive security audits. Given that Harness can execute model-generated code, load third-party plugins, and interact with network resources, its capabilities pose inherent risks. Incorrect coding, misconfigurations, or unverified plugins might compromise the host, potentially altering or erasing files, or exposing sensitive data and credentials.

While implementing sandboxing, approval prompts, and permission controls can help mitigate these risks, they do not guarantee full security or protection against damage, even when correctly set up.