What is AG-UI?
AG-UI (Agent-User Interaction protocol) defines how an agent backend talks to the app a person is using. The app sends one RunAgentInput (messages, the tools the front end offers, shared state and context) to an HTTP endpoint, and the agent streams back typed events: 31 of them in eight families covering run lifecycle, text, tool calls, state snapshots and JSON Patch deltas, reasoning, progress activity, subagents and custom data. Transports are HTTP with Server-Sent Events or a binary Protobuf binding.
It sits beside the other two agent protocols rather than competing with them. MCP connects an agent to tools and data, A2A connects agents to each other, and AG-UI connects an agent to a user; one agent often speaks all three, and AG-UI middleware can front MCP and A2A agents. Generative UI specs such as Google's A2UI and MCP Apps describe what to draw, and AG-UI carries those payloads.
Version 1.0 shipped on 17 September 2026 with a normative specification and JSON Schema, from which the official TypeScript, Python and .NET SDKs are generated; Kotlin, Go, Java, Dart, Rust, Ruby, C++ and Swift SDKs come from the community. CopilotKit created the protocol and maintains it in the open under MIT.
Is AG-UI worth it?
AG-UI solves the part of agent work every team rebuilds badly: getting streamed text, tool calls, shared state and approval prompts from a backend agent into a browser without inventing a custom event format per framework. The adoption numbers say it has largely won that argument. @ag-ui/core was downloaded about 7.6 million times on npm in the week to 19 November 2026, and the Python ag-ui-protocol package about 2.3 million times; the repository has about 20,800 GitHub stars.
Framework support is the reason. Microsoft Agent Framework, Google ADK, AWS Strands, Mastra, Pydantic AI, Agno, LlamaIndex and AG2 build it in; LangChain and CrewAI are listed partners; Claude Agent SDK and OpenAI Agents integrations live in the repository; and Amazon Bedrock AgentCore documents AG-UI for its runtime. Swapping the agent framework behind a CopilotKit or terminal front end becomes a configuration change rather than a rewrite.
The 1.0 release on 17 September 2026 moved behaviour out of the TypeScript client and into a written spec with MUST and SHOULD rules. Runs now report an outcome (success, interrupt or cancelled), which turns human approval into a clean pause-and-resume; THINKING events became REASONING events with encrypted round-trips; and tool results can carry images and documents. Old and new versions interoperate in both directions, so a 0.x agent keeps working with a 1.0 client and the reverse.
The catch: the protocol has no authentication or authorisation of its own. Microsoft's AG-UI security guide says so directly, warns against exposing an AG-UI server to browser JavaScript or mobile apps, and lists the messages list and state as the main routes for prompt injection, because a client can send system messages, tools or state of its choosing. Thread IDs are not credentials. In practice you put a trusted backend in front that builds the protocol messages and passes through only the user's text. The release cadence is also fast, with tagged releases every few days in November, and 318 issues are open.
Our take: if your agent needs a real user interface, AG-UI is the sensible default in the way MCP is for tools; budget a day for the security plumbing the spec deliberately leaves to you.
Where it earns its place
A front-end engineer at a SaaS company adding an in-app assistant
Show streamed answers, tool progress and an approve-or-reject step for risky actions
Run `npx create-ag-ui-app@latest`, choose the backend framework and CopilotKit on Next.js, then map interrupt outcomes to a confirmation dialog. The protocol is free; cost is the model calls and hosting of the agent.
A platform team at a bank running agents on two frameworks
Keep one chat front end while the backend moves from LangGraph to Microsoft Agent Framework
Both emit AG-UI events, so the client stays the same. Put a backend-for-frontend service in front that authenticates users, builds RunAgentInput itself and forwards only user text, as Microsoft's security guide recommends.
A developer building an internal Slack bot on an existing agent
Reuse the web agent in Slack and Teams without a second integration
CopilotKit's channels SDK consumes the same AG-UI stream as the web client, so the agent code does not change; only the rendering of tool calls and state differs per channel.
Before you commit
- Setup
- npm install @ag-ui/client or pip install ag-ui-protocol; a minimal Python agent and TypeScript client is about 30 lines each, or `npx create-ag-ui-app@latest` scaffolds a CopilotKit app wired to a framework (the templates call OpenAI).
- Integrations
- Microsoft Agent Framework, Google ADK, AWS Strands, Mastra, Pydantic AI, Agno, LlamaIndex, AG2, LangChain, CrewAI, Claude Agent SDK, OpenAI Agents, Spring AI and more; clients for React, Angular, Vue, React Native, Slack, Teams and terminals; MCP and A2A middleware.
- Data and privacy
- A protocol, not a service: events flow only between your app and your agent. No authentication is defined, so endpoints need your own auth and a trusted backend in front; MIT licence.
- Cost at scale
- No licence or usage fee; the cost is your agent hosting, model tokens and long-lived streaming connections at your load balancer.
- Limits
- Python SDK needs 3.9+ and the TypeScript client Node.js 20+; 1.0 drops unknown events with a warning and treats malformed known events as fatal, so mismatched SDK versions surface as runtime errors.
What’s good
- Built into most major agent frameworks, including Microsoft, Google and AWS
- Frozen 1.0 spec and JSON Schema with generated TypeScript, Python and .NET SDKs
- Interrupt outcomes make human approval a clean pause and resume
- Shared state with JSON Patch deltas keeps app and agent in sync
- Live demos for every integration in the Dojo
What’s not
- No built-in authentication; endpoints must not face browsers directly
- Messages and state are open prompt-injection routes from an untrusted client
- Kotlin, Go, Rust and other SDKs are community-maintained, not generated from the schema
- Fast release cadence and 318 open issues
Best for
Teams putting an agent built on LangGraph, ADK, Mastra, CrewAI or similar behind a web, mobile or chat interface who want one event format for all of them.
AG-UI alternatives
AG-UI: common questions
- Is AG-UI worth it?
- AG-UI has become the default wire format between agent backends and front ends, and its 1.0 spec makes it safe to build on, as long as you supply the authentication it leaves out. We rate it 4.3 out of 5.
- How much does AG-UI cost?
- Free and open source under the MIT licence: protocol, SDKs, integrations and the Dojo demo app. You pay for your own agent hosting and model calls.
- What are the downsides of AG-UI?
- No built-in authentication; endpoints must not face browsers directly; messages and state are open prompt-injection routes from an untrusted client; kotlin, Go, Rust and other SDKs are community-maintained, not generated from the schema; fast release cadence and 318 open issues.
- Who is AG-UI best for?
- Teams putting an agent built on LangGraph, ADK, Mastra, CrewAI or similar behind a web, mobile or chat interface who want one event format for all of them.
- How much does AG-UI cost for a team or in production?
- No licence or usage fee; the cost is your agent hosting, model tokens and long-lived streaming connections at your load balancer.
- Is AG-UI safe for company data?
- A protocol, not a service: events flow only between your app and your agent. No authentication is defined, so endpoints need your own auth and a trusted backend in front; MIT licence.
- What are the best AG-UI alternatives?
- The closest alternatives we have reviewed are MCP Python SDK, Mastra, LangGraph and Google Agent Development Kit (ADK).
Answers checked Nov 20, 2026.