Rank #06 · Weekly · Infrastructure
AG-UI: The Agent-User Interaction Protocol
Give any agent a typed, bidirectional event stream so a web app, mobile client, terminal, or chat platform can render the same run.
What it does
AG-UI is the Agent-User Interaction Protocol: an open, lightweight, event-based protocol that standardizes how AI agents connect to user-facing applications. GitHub description: "AG-UI: the Agent-User Interaction Protocol. Bring Agents into Frontend Applications." Site: ag-ui.com. Docs: docs.ag-ui.com. Spec: docs.ag-ui.com/spec/1.0. Examples: dojo.ag-ui.com.
README and official docs (as verified):
- Agents stream tokens, call tools mid-response, update shared state, pause for a human, and spawn subagents — a poor fit for a one-shot request/response API. AG-UI replaces per-framework custom wiring with one bidirectional event stream: the agent emits typed events, and any client can render them. The same agent can power a web app, a mobile app, a terminal, or a chat platform.
- A client starts a run by sending a single
RunAgentInput(conversation, frontend-offered tools, shared state, and context). The agent answers with a stream of typed events. AG-UI 1.0 defines 31 event types in eight families: runs and steps, text messages, tool calls, state (STATE_SNAPSHOT, JSON PatchSTATE_DELTA,MESSAGES_SNAPSHOT), reasoning, activity, subagents, and passthrough (RAW,CUSTOM). Only the run lifecycle is required, so a minimal agent can start small. - The protocol is transport-agnostic. The 1.0 spec defines HTTP + SSE and a binary HTTP + Protobuf binding; middleware can adapt events from other sources.
- AG-UI 1.0 is described as a stable, frozen specification backed by a JSON Schema. The TypeScript, Python, and .NET SDKs are generated from that schema. The README states that 1.0 is backwards compatible in both directions with 0.x agents and clients.
- Scaffold a starter with
npx create-ag-ui-app@latest(a CopilotKit Next.js frontend, or a terminal client, wired to an agent framework you pick). First-party packages include@ag-ui/client/@ag-ui/coreon npm,ag-ui-protocolon PyPI, andAGUI.Clientfor .NET. Community SDKs for other languages are listed in the README. - Documented building blocks include streaming chat, generative UI, shared state, frontend tools, human-in-the-loop interrupts, reasoning display, subagents, and multimodal input. Live demos are collected in the AG-UI Dojo.
- AG-UI is positioned as complementary to MCP (agent ↔ tools and data) and A2A (agent ↔ agent). It is also distinct from generative UI specs such as A2UI and MCP Apps: those describe what UI an agent wants to show; AG-UI is the runtime protocol that can carry those payloads along with everything else.
- The README lists integrations across agent frameworks (including LangChain, CrewAI, Microsoft Agent Framework, Google ADK, AWS Strands Agents, Mastra, Pydantic AI, LlamaIndex, and others in the docs). Treat partner and "adopted by" wording as the project's claim, not a TidyHow census. AG-UI was created by CopilotKit and is developed in the open.
License: MIT (LICENSE file and GitHub SPDX agree).
Who it's for
- Application teams that want to connect an agent backend to a user-facing app (web, mobile, terminal, or chat) without inventing a private event schema.
- Framework or platform authors adding a standard agent-to-UI transport beside MCP and A2A.
- Client builders who need to render streaming chat, shared state, tool calls, or human-in-the-loop interrupts from any AG-UI-compatible agent.
Less ideal if you only need a single-shot chat completion with no streaming, shared state, or mid-run user input, or if you want a hosted agent product rather than a protocol and SDKs to implement.
Stars / language / license
20,638 stars · TypeScript · MIT
Data verified: 2026-11-20 09:04 Asia/Shanghai (official Daily Top10 gate; language/license confirmed via GitHub API get_repository + README + LICENSE + official docs)
Trending note: +3,455/wk (weekly Trending window; selection captured 09:04 Asia/Shanghai)
Category tags
Infrastructure · agent-protocol · event-stream · TypeScript · MIT · generative-UI · human-in-the-loop · CopilotKit
How it differs from similar projects
Compared with a custom REST, GraphQL, or ad-hoc SSE API. Those can stream tokens, but every agent framework tends to invent its own event shapes for tools, state, and interrupts. AG-UI's pitch is a shared event vocabulary and generated SDKs so the client does not change when you swap the agent.
Compared with MCP. MCP connects agents to tools, workflows, and data sources. AG-UI connects agents to user-facing applications. The docs describe them as complementary layers; a single agent may use both. (On today's board, Firecrawl #01 and chrome-devtools-mcp #07 also expose MCP surfaces — different problem domains from AG-UI's agent-to-UI protocol.)
Compared with A2A. A2A is for agent-to-agent coordination. AG-UI is for agent-to-user interaction.
Compared with A2UI, MCP Apps, and similar generative UI specs. Those specify UI payloads. AG-UI is the runtime protocol that can transport them, plus chat, state, tools, and interrupts.
Caveats. Integration depth varies by framework; check the official list and Dojo demos rather than assuming every listed partner exposes every event family. Some example clients use hosted services (for example CopilotKit Intelligence in chat-platform samples) that are not part of the protocol. No independent adoption, latency, or compatibility audit is claimed here.