ag-ui-protocol/ag-ui

Overview

AG-UI is an open, event-based protocol that standardizes how AI agents connect to user-facing applications. Agent backends emit events matching roughly 16 standard event types and accept a small set of AG-UI-compatible inputs, while a middleware layer bridges transports such as SSE, WebSockets and webhooks and tolerates loose event-format matching. A reference HTTP implementation and default connector are included.

It targets teams building in-app agent experiences: streaming chat, bidirectional state sync, generative UI and structured messages, real-time context enrichment, frontend tool calls and human-in-the-loop approval. In the agent stack, the project positions AG-UI as the user-facing layer alongside MCP (tools) and A2A (agent-to-agent communication).

Getting started

You need a Node.js environment with npx available. Create a starter application from the repository root or any working directory:

npx create-ag-ui-app my-agent-app

This scaffolds an application named my-agent-app; the README does not document the generated structure or the run command, so check the scaffold output or the application quickstart for the next step. To see working behavior without writing code, open the AG-UI Dojo, which demonstrates each core building block in examples of roughly 50–200 lines and hosts demos for the supported framework integrations.

Integrations and SDKs

First-party support covers the built-in agent, LangChain/LangGraph, CrewAI, Microsoft Agent Framework, Google ADK, Google Antigravity, AWS Strands Agents, Mastra, Pydantic AI, Agno, LlamaIndex and AG2; AWS Bedrock Agents and the OpenAI Agents SDK are in progress. Community integrations include Claude Agent SDK, Claude Managed Agents, Langroid, Omnara and Cloudflare Agents. SDKs exist for Kotlin, Go, Dart, Java, Rust, Ruby, C++, .NET and Nim, with Flowise and Langflow in progress. Client-side support includes CopilotKit, terminal/agent clients, Slack and Teams via the Channels SDK, React Native and OpenBot.

Constraints

The protocol is MIT-licensed and designed to be transport-agnostic, so it does not force a particular event transport or framework. Compatibility depends on using an AG-UI connector or adapter for the chosen backend; several integrations are marked in progress and are not yet usable. The repository's quick start stops at project scaffolding, so the full flow to a running app requires the linked documentation. Human-in-the-loop and frontend tool features require the frontend side to handle the corresponding events.

HotRepo's take

AG-UI reached 8th place on HotRepo's all-language GitHub daily ranking, recorded on 2026-11-14 with 17,183 stars and 9 stars gained that day. Its Show HN thread from 2025-05-13 drew 36 points and 5 comments, and the discussion shows both the appeal and the confusion the project has to manage: one commenter (dhorthy) framed it as a client-agnostic way to display what an agent is doing — working, thinking, calling tools, errors, needing input or approval — while another (aatd86) asked what it is actually for and whether it is an accessibility layer or a new UI paradigm. That gap suggests the protocol's value depends on concrete frontend use cases, which is why the Dojo examples matter more than the specification alone.