deepseek-harness-genui

Think about asking an AI to help you block out three hours of focused writing time. Describing every constraint and preference in plain text always feels a bit awkward. DeepSeek Harness GenUI exists for exactly these moments: it lets an Agent create a dedicated React interface for the task at hand—whether that's picking calendar slots, exploring how photosynthesis reacts to different variables, or tracing the flow of a real codebase. What was hard to explain in words becomes something you can see and interact with.

The plugin is code-first. Instead of relying on a component-tree DSL or a prebuilt widget catalog, the Coding Agent writes ordinary React + TypeScript, and the plugin builds, checks, and runs it in a sandbox. Whatever the user selects, enters, or changes is saved as semantic task state; the next Agent turn can read that state and continue without making the user repeat anything. A failed update never replaces the last working version, so iteration stays safe.

It doesn't stop at rendering, either. Generated apps can only call declared, credential-free public HTTPS routes or approved tools; before a connected app opens, Harness shows the full access list for a single task-scoped decision, and undeclared calls are simply blocked. Design is managed through DESIGN.md, letting you pick a built-in profile like Material 3, Apple Human Interface, or shadcn/ui, or import your own. The same app can sit inline, open in Canvas, go fullscreen, or appear as a localhost link—all surfaces share the same task state.

This plugin is for DeepSeek Harness users who need purpose-built interfaces for tasks that aren't known in advance: free-form simulations, spatial tools, connected workflows, or anything where the user's complex choices need to become state for the next step. If a compact card or simple form assembled from a known catalog is enough, lightweight dsh-ui component renderers are still the right choice. But when the structure and interactions can't be predicted ahead of time, GenUI's code-first approach gives the Agent the freedom to build exactly what the task requires.