A2UI is an open-source project that provides a declarative JSON format and renderer implementations enabling AI agents to generate rich, interactive user interfaces. The project addresses a fundamental challenge in agentic AI: while language models excel at producing text and code, they struggle to create safe, expressive interfaces that work across different platforms and trust boundaries. A2UI solves this by allowing agents to describe UI intent in a standardized format that client applications then render using their native component libraries, whether Flutter, Angular, Lit, React, or other frameworks.
The core architecture separates UI generation from UI execution. An agent generates an A2UI Response as a JSON payload describing component composition and properties. This message travels to a client application, where an A2UI Renderer parses the JSON and maps abstract component types to concrete implementations in the client's codebase. This separation ensures that agent-generated UIs are safe like data—since they cannot execute arbitrary code—while remaining expressive like code through a rich component vocabulary.
Security is foundational to A2UI's design philosophy. Rather than allowing agents to execute code, client applications maintain a catalog of trusted, pre-approved UI components that agents can request. This places security firmly in the developer's hands, enabling strict sandboxing policies and trust ladders within custom component logic. The format is also designed to be LLM-friendly, using a flat list of components with ID references that models can generate incrementally, enabling progressive rendering and responsive user experiences. Agents can efficiently make incremental changes to the UI as conversations progress.
The project is framework-agnostic and portable by design. The same A2UI JSON payload from an agent can be rendered across multiple different clients built on different frameworks. An open registry pattern allows developers to map server-side types to custom client implementations, from native mobile widgets to React components, including secure iframe containers for legacy content.
According to GitGenius activity tracking, the repository shows strong engagement with 568 tracked issues and pull requests. The median response latency is 0.0 hours with a mean of 52.6 hours, indicating active maintenance. The most frequently labeled issues are P2 priority items (259), feature and enhancement requests (198), and sprint-ready items (121). Top contributors include jacobsimionato with 999 events, gspencergoog with 585 events, and selamw1 with 460 events. The project shares overlapping contributors with github/gh-aw, solo-io/gloo, and flutter/flutter, suggesting integration with broader ecosystem initiatives.
A2UI's current production release is v0.9.1, a stable patch in the v0.9 protocol family, while v1.0 is a release candidate. The project is in early-stage public preview, with the specification and implementations functional but still evolving. The team is actively gathering feedback and soliciting contributions, particularly for client renderers.
Use cases span dynamic data collection where agents generate context-specific forms, remote sub-agents that return UI payloads for rendering in chat windows, and adaptive workflows where enterprise agents generate approval dashboards or data visualizations on demand. The project integrates with the A2A Protocol and AG-UI for transport, works with any JSON-capable LLM including Gemini, and currently supports web and Flutter host frameworks.
Getting started options include a quickstart restaurant finder demo using Gemini and Lit renderer, integration with AG-UI and CopilotKit for any agent framework, a visual A2UI Composer for generating JSON without installation, and A2UI Theater for stepping through pre-built streaming scenarios. The roadmap includes specification stabilization toward v1.0, additional renderers for React, Jetpack Compose, and iOS SwiftUI, support for REST transport, and integration with more agent frameworks like Genkit and LangGraph. The project is Apache 2.0 licensed and actively welcomes community contributions.
