DeepSeek Harness GenUI plugin
A DeepSeek Harness plugin where the agent builds a task-specific React app and reads the user's saved choices in the next turn.
- 124 ★ on the repository
- Last commit
Extended access
Why this level
- The agent writes and runs interface code, and declared tools and public HTTPS routes are reached through a broker.
Install
Manual install
dsh plugin --profile web add dsh-plugin-genuiThe single install command from the README, then `dsh --profile web`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent writes and runs interface code, and declared tools and public HTTPS routes are reached through a broker.
This is third-party code. Review the repository files before installing.
What it does
DeepSeek Harness GenUI helps an agent build a focused interface when a task is awkward in text. The coding agent writes plain React and TypeScript, not a component-tree description. The user saves values in the interface: chosen slots, form answers, a draft, progress. The agent's next turn reads them and continues the task without re-entry. README examples: picking calendar slots, an interactive photosynthesis model with four controls, a code-path explorer. The app opens inline, in Canvas, fullscreen or on localhost. Calls to connected tools and public HTTPS routes must be declared up front, and undeclared calls are blocked. A version that fails the build or source checks does not replace the working one.
Who it is for. DeepSeek Harness users on the Web profile who need interactive planners, simulators and code explorers inside the conversation.
Good fit when
- You plan a Saturday route, move visit times and make the garden optional.
- A physics model such as a double slit is explored with sliders for wavelength and distance.
- An explanation of a CLI turns into an explorer of files, functions and branches.
- The results of a choice must carry into the next agent turn without retelling.
Not a fit when
- Plain questions, rewriting, summaries and short lists are enough: keep them as prose.
- You need predictable cards, tables and forms: schema renderers fit better.
- You use a TUI or headless profile: they are not supported.
- You need a cross-client UI protocol: the plugin is tied to DeepSeek Harness.
Example request
Build an interface for a Saturday plan: a museum, a riverside garden and dinner. I want to change the times and drop the garden, then continue from my choice.Limitations
Node.js `^22.19.0 || ^24.0.0` and a supported DeepSeek Harness Web profile are required. Version 0.15.2 supports Harness `0.2.1-rc.2` and `0.2.1-rc.3`. TUI and headless profiles do not work. Generated code runs in an isolated sandbox, but the boundary does not make malicious code safe with data it was given access to: keep secrets out of app state. Temporary links and grants last 7 days, and task state is kept 7 days after its last update.
How to disable. Disable the plugin in DeepSeek Harness settings (Settings, Plugins) or remove it with `dsh plugin` for the `web` profile. Task state expires after 7 days.
FAQ
How do I install the plugin?
Run `dsh plugin --profile web add dsh-plugin-genui`, then `dsh --profile web`. Node.js `^22.19.0 || ^24.0.0` is required; Chrome is not needed for the plugin.
How does it differ from schema-based UI renderers?
Schema renderers suit predictable cards and forms. The plugin writes plain React and suits simulators and spatial tools, but is tied to DeepSeek Harness.
How does the agent learn what the user chose?
The user saves values in the interface and the agent's next turn reads them. Test it by asking: "What did I just choose in the interface?"
Can I pick a visual style?
Yes. Settings offer automatic selection, `material-3`, `apple-human-interface`, `shadcn-ui` or an imported custom `DESIGN.md`.
How long is state kept?
Task state is kept for 7 days after its last update. Temporary links and grants also expire after 7 days.
Discussion
No questions or reviews yet.
Related
PluginsAnnotation toolbar for React pages: click an element to hand the agent its selector and context, plus skills for setup and an autonomous UI critique
A CLI that lets an agent inspect a running React app: component tree, props, state, hooks and render profiling
An interface generator that turns a model's description into a validated React component built on the Gravity UI library
Bitovi's plugin marketplace: specs and feature delivery, Figma and React, Playwright QA, tRPC and Prisma
A skill that picks and installs beUI animated React components through the public MCP or the shadcn registry instead of hand-built widgets
A CLI that sets up a canvas where an AI agent builds dashboards and reports as React artifacts, revised block by block
README in short
The README covers installation, the core state-saving loop, display surfaces, `DESIGN.md` configuration and the safety model. The project's research basis is the EvoGenUI-Bench paper.