What if an AI assistant could return a real form or interactive card instead of another wall of text? Google’s A2UI protocol is designed for exactly that—and you do not need to let a model write arbitrary HTML.
On October 29, 2026, Google’s developer team published a practical guide to combining A2UI with CopilotKit. Here is the useful part for developers.
What A2UI actually does
A2UI (Agent-to-UI) lets an agent describe an interface as structured data. Your app renders that description using components you have already approved. Think cards, buttons, forms, and choice pickers—not untrusted JavaScript pasted into your page.
CopilotKit connects that agent to the frontend. The AG-UI event protocol carries updates and user actions between the two.
- A2UI: describes which components to show and their properties.
- Component catalog: limits the agent to UI elements your app permits.
- CopilotKit / AG-UI: handles the agent-to-app interaction and updates.
Try the official starter
Google’s October 29 guide gives this scaffold command. It is copied from the official instructions, not a command we have tested locally:
npx copilotkit@latest create --name my-app --framework a2uiReview the generated project, dependency versions, and required API configuration before running it with credentials. The starter is for experimenting with an A2UI-enabled agent app, not a production-ready deployment.
Why this is more useful than generated HTML
Suppose a travel assistant finds three flights. Instead of describing them in paragraphs, it can ask your frontend to render approved flight cards and a cabin selector. The user changes an option, and the selection travels back to the agent.
That is the practical win: interactive results inside the conversation, while your app retains control over what components exist.
The important limitation
A2UI is still evolving. Its official project describes v0.9.1 as the current production release and v1.0 as a release candidate. Check the version before building around a specific schema.
Declarative UI also does not make every agent action safe. Validate permissions and sensitive operations on your server, and never assume a generated form is trustworthy just because its components are approved.
Takeaway: if you are building an AI assistant that needs forms, cards, or multi-step interactions, try the starter and define a small, trusted component catalog first.