AG-UI (Agent-User Interaction Protocol)

What it is

AG-UI — short for the Agent–User Interaction Protocol — is an open, event-based standard for connecting agents to user-facing applications^[1]. It is defined as a general-purpose, bi-directional connection between a user-facing application and any agentic backend, rather than a link to one particular chat product or framework^[1]. The protocol therefore sits on the boundary where the agent runtime meets the interface a person is actually looking at, and describes what may cross that boundary.

How it works

AG-UI standardises how agent state, UI intents and user interactions flow between the agent runtime and the front end^[1]. "Event-based" means the connection carries a stream of discrete messages rather than a single request and a single reply: each update from the agent arrives as an event that the front end can act on as it comes in. In practice the protocol provides live token and event streaming for multi-turn sessions, and a session can be cancelled and resumed^[1]. Token streaming matters because a model's answer arrives gradually, and showing it as it is produced is what keeps an agent from looking frozen. Cancel and resume matter because real sessions get interrupted — a user stops a long answer, or comes back to a conversation later — and the connection has to survive that.

Alongside text, AG-UI supports typed attachments and real-time media, including files, images, audio and transcripts^[1]. Typed payloads carry a declared type, so each side knows what it is sending and receiving instead of inferring structure from an unlabelled blob.

Where it sits

The AG-UI documentation places the protocol alongside MCP, which originated at Anthropic, and A2A, which originated at Google^[1]. It also draws a line between AG-UI and A2UI: A2UI is about agents delivering UI widgets, whereas AG-UI is about connecting an agentic front end to any agentic backend^[1]. The two distinctions are different in kind — one is about which protocol covers which part of an agent system, the other about what the agent puts on the screen versus how the front end talks to the agent at all.

What ships with it

The project ships SDKs for TypeScript, Python and .NET, together with a formal specification^[1]. Having both a written specification and SDKs in several languages is what lets different teams implement the same wire behaviour on either side of the connection.

Status and direction

The project publishes proposals for generative UI and meta events, and a migration guide to version 1.0^[1]. A proposals process is how a specification project floats design changes in public before they become part of the standard; a migration guide is what lets implementations written against earlier behaviour be brought forward.

Among the protocols the documentation lists, AG-UI is the one aimed at the user-facing end: MCP and A2A stand beside it as neighbours, and A2UI as the widget-level counterpart^[1].