Introduction
An Angular schematic and library that turns the features of your app into interactive widgets in an AI chat.
You have an Angular app. You want its features to appear as interactive widgets in an AI chat: a poll that the user can vote on, a chart, a form. Your app serves them, and your tools drive them.
ng-mcp-ui does this with one schematic.
ng add ng-mcp-ui --example=demoThe schematic mounts an MCP server into the Angular SSR
server.ts file of your app. It also scaffolds an example tool and widget, and it adds a dev
tunnel. Thus you can connect a real host, for example Claude or ChatGPT, and see your changes
immediately.
What the package contains
| Import | Purpose |
|---|---|
ng-mcp-ui/server |
The framework-neutral MCP server: McpServer, the mountable Express router, the content and FileRef helpers, the view shells, and auth |
ng-mcp-ui/web |
The Angular host bridge: provideMcpUi, bootstrapWidget, the inject* signal API, the [dataLlm] directive and the mcpAsset pipe |
ng-mcp-ui/testing |
The MockAdaptor and provideMockMcpUi harness for tests and Storybook |
ng-mcp-ui/tunnel |
A marker for a cloudflared dev tunnel. It is a skeleton today |
The same package also contains the Angular schematics (ng-add, view,
tool and example) and the ng-mcp-ui:build-widgets builder.
The pack step embeds them under dist/schematics/.
What you can do
- Retrofit an Angular app with
ng add. CI proves this on Angular v20, v21 and v22. - Write one widget for two host runtimes. One
Adaptorinterface covers Claude and ChatGPT. - Move typed data from a tool to a view. The server infers your Zod schemas end to end, through
typeof serverintoinjectAppHelpers. - Call a server tool from the view, keep view state on the host, and show in-view content to the
model with
[dataLlm]. - Match the theme, the display mode and the safe-area insets of the host.
- Connect a real host through a dev tunnel that needs no auth.
- Unit-test a widget with
MockAdaptorandprovideMockMcpUi.
Host support
One Adaptor interface covers the two runtimes: the OpenAI Apps SDK (window.openai, which
ChatGPT uses) and the open MCP-Apps postMessage specification
(@modelcontextprotocol/ext-apps, which Claude and other MCP-Apps hosts use). Therefore your widget
code is the same for each host.
Live-host tests are signed off on Claude and on ChatGPT. A machine verifies the render, and a person confirms the interactive tool-call rows. Gemini is not supported.
For the behavior of each function on each host, see host support.
Where to go next
- Quickstart gives you a retrofitted app in about two minutes.
- How it works gives you the architecture, and it tells you why the package uses SSR for the server and not for the render.
- Schematics gives you each generator and its options.