MCP handles tools, A2A handles agents, AG-UI handles users

What it does

AG-UI defines a common language for agent backends and user interfaces. It specifies roughly sixteen standard event types—covering streaming messages, state changes, and tool calls—that any agent can emit and any frontend can consume. A transport-agnostic middleware layer handles the actual plumbing across SSE, WebSockets, or webhooks, so the protocol stays loose enough to bridge diverse frameworks without forcing strict format lockstep.

The interesting bit

The project explicitly carves out the “agent-to-human” layer in the protocol stack, leaving MCP for tools and A2A for agent-to-agent chatter. That positioning is either prescient or presumptuous, but the breadth of buy-in is hard to ignore: LangGraph, CrewAI, Google ADK, Mastra, Pydantic AI, and several others already ship first- or third-party integrations, plus community SDKs in Rust, Go, Kotlin, and more.

Key highlights

• Transport-agnostic by design: works over SSE, WebSockets, or webhooks
• ~16 standard event types for streaming chat, state sync, generative UI, and human-in-the-loop flows
• Reference HTTP implementation and default connector included
• Integrations span major frameworks (LangGraph, CrewAI, LlamaIndex, AG2, AWS Strands, etc.)
• Community SDKs in Kotlin, Golang, Dart, Java, Rust, Ruby, and C++

Caveats

• Several integrations are still marked “In Progress”: AWS Bedrock Agents, OpenAI Agent SDK, Cloudflare Agents, React Native, .NET, Nim, Flowise, and Langflow
• The “loose event format matching” that enables broad interoperability also means implementers must pay attention to edge-case behavior across frameworks

Verdict Worth a look if you are building a frontend that needs to talk to multiple agent backends without writing bespoke adapters for each. Skip it if your agent interaction is a one-off chat stream that does not need standardized state synchronization or generative UI contracts.