Agentic GUI Protocols: Google's A2UI and CopilotKit's AG-UI

Forum topic · ✨步子哥 · 2026-03-29

Summary

This forum post surveys the emerging Agentic GUI protocol ecosystem, centered on two open-source projects: Google's A2UI (Agent-to-User Interface) and CopilotKit's AG-UI (Agent-User Interaction Protocol). A2UI is a declarative UI protocol where agents describe interfaces using safe JSON with a whitelisted set of ~18 Material-style components, rendered natively via official renderers for Lit, Angular, Flutter, and React, avoiding XSS risks from raw HTML. AG-UI is a lightweight event-driven protocol using SSE streaming and 16 standard event types to synchronize agent backends (LangGraph, CrewAI, ADK) with frontend UI state in real time. The two are complementary: AG-UI handles communication while A2UI defines UI content format. Related projects include Google's A2A for agent-to-agent communication and Anthropic's MCP for tool/context integration. The post includes setup guidance, noting CopilotKit is the official AG-UI implementation adopted by LangChain, AWS, and Microsoft.

The main open-source projects in the Agentic GUI protocol space (also called Agentic UI or Generative UI protocols) are Google's A2UI and the CopilotKit team's AG-UI. These two projects form the core of the current ecosystem and address the same problem space ("AI2UI" is essentially an alternate name or typo for A2UI).

Both protocols solve a core challenge: allowing AI agents to safely generate or interact with rich UI across platforms — no longer limited to plain text output — while avoiding the security risks (such as XSS) of directly executing HTML/code, and supporting native rendering on web, mobile, and desktop.

1. A2UI (Agent-to-User Interface) — Google's declarative UI protocol

GitHub: https://github.com/google/A2UI (Apache 2.0 license, actively developed)
Website: https://a2ui.org/
Key features:
Agents describe UI intent using safe, declarative JSON components from a whitelisted component library (~18 base components such as rows, columns, buttons, and forms, in Material style).
Clients render with native components — official renderers exist for Lit, Angular, Flutter, React, and more — ensuring cross-platform consistency and safety.
Designed specifically for "agents generating dynamic interactive interfaces," and already integrated into Google Chat, Vertex AI, and other products.
Use cases: when an agent needs to generate forms, dashboards, rich chat interfaces, and other complex UI in real time.
Community support: frameworks like CopilotKit and LangGraph already provide integration examples.

2. AG-UI (Agent-User Interaction Protocol) — CopilotKit's real-time event protocol

GitHub: https://github.com/ag-ui-protocol/ag-ui (MIT license)
Website/docs: https://www.copilotkit.ai/ag-ui
Key features:
A lightweight, event-driven protocol (SSE streaming plus 16 standard event types such as TEXT_MESSAGE_CONTENT, TOOL_CALL_START, UI_UPDATE, etc.).
Handles real-time bidirectional synchronization between the agent backend and frontend app, mapping output from agent frameworks (LangGraph, CrewAI, ADK, etc.) directly to UI state.
Highly complementary to A2UI: AG-UI handles "communication," A2UI handles "UI content format" — the two are often used together.
Use cases: building full-stack agentic applications that need real-time streaming UI updates, tool call feedback, and user intervention.
Strongest ecosystem: CopilotKit ( https://github.com/CopilotKit/CopilotKit ) is the official implementation, adopted by LangChain, AWS, Microsoft, and others, offering React/Angular SDKs and many Generative UI examples.

Related projects (part of the full stack, but not GUI protocols themselves)

A2A (Agent2Agent Protocol): Google-led agent-to-agent communication protocol ( https://github.com/a2aproject/A2A ) for multi-agent collaboration; can be combined with AG-UI/A2UI to push results to the frontend.
MCP (Model Context Protocol): Anthropic-led tool/context protocol ( https://modelcontextprotocol.io/ ); some extensions support MCP Apps (interactive UI), but it is primarily for tool calling rather than pure GUI.

Getting started

To try A2UI directly: clone the Google repository and run the demo with Google ADK (Agent Development Kit).
To quickly build a full-stack app: use CopilotKit + AG-UI + A2UI (official starter repos exist, e.g., with-agent-spec).

As of March 2026, the Agentic GUI ecosystem is still evolving rapidly, and these two projects are the de facto standards that nearly all major agent frameworks are aligning with.