Introduction

AG-UI is an open, lightweight, event-based protocol for connecting AI agents to user-facing applications. It is described as a general-purpose, bi-directional connection between a user-facing application and an agentic backend, intended to help application developers build reliable, debuggable, user-friendly agentic features while avoiding complex ad-hoc wiring.

Its documented building blocks include live token and event streaming for multi-turn sessions with cancel and resume, typed attachments and real-time media for files, images, audio, and transcripts, and stable typed components for rendering model output under app control. The protocol also supports a small declarative language for generative UI, where agents propose trees and constraints and the application validates and mounts them. Other listed capabilities include shared state with streamed event-sourced diffs and conflict resolution, visualization of intermediate reasoning from traces and tool events, typed handoffs for frontend tool calls, backend tool output rendering, human-in-the-loop interrupts, nested sub-agents with scoped state, real-time user steering, tool result and log streaming, and custom events.

The documentation frames AG-UI as one of three open agentic protocols, alongside MCP and A2A, and says it is the Agent–User Interaction standard. It is built on HTTP and WebSockets. Supported integrations shown on the page include direct-to-LLM, agent framework partnerships, first-party agent frameworks, and community agent frameworks, and the page states that AG-UI was born from CopilotKit’s initial partnership with LangGraph and CrewAI. The documentation also points to SDKs in TypeScript, Python, and .NET, plus a Discord community and an AG-UI repository.

The page does not state pricing or licensing details.