AGenUI/AGenUI

AGenUI is an open-source SDK that turns streaming output from a large language model into native, interactive mobile interfaces on all three major mobile platforms at once. When an LLM agent decides what a screen should look like, it emits a structured message describing the UI (text, buttons, forms, lists, images, video, etc.). AGenUI parses that message as it arrives — piece by piece, like a live feed — and draws the matching native UI components on the device, so the interface appears and updates while the model is still generating it.

It's built on the A2UI v0.9 protocol, which Google open-sourced as a standard way for LLM agents to describe UI. AGenUI fully implements that spec and adds its own extensions.

How it's engineered: a shared C++ core engine handles the heavy lifting common to all platforms — streaming data parsing, building a virtual component tree, caching render data, resolving theme/style, and diffing changes. Below that, three platform-specific rendering engines (iOS, Android, HarmonyOS) receive the parsed component descriptions from the C++ layer and draw them using each system's native UI capabilities. This delivers smooth, high-refresh rendering (120 fps in core scenarios like scrolling).

Key capabilities:

• 22 built-in components — 18 from the A2UI spec (Text, Image, Button, Card, List, Tabs, Slider, CheckBox, Modal, etc.) plus 4 SDK extensions (Table, Carousel, Web, RichText). You can also register your own custom native components at runtime.
• Function Call integration — developers can register client-side tools the LLM can invoke on demand. v1.7.0 adds a resolveCalls built-in that recursively executes nested function calls in JSON and replaces each call node with its result.
• Design Token & theming — a unified token system and light/dark mode across all three platforms.
• Catalog file (agenui_catalog.json) — a self-contained, freestanding JSON Schema (conforming to A2UI v0.9) describing all 25 components and 14 functions. Agents use it as a constraint so the LLM emits strictly conformant UI descriptions.
• A2UI Generation Skill (skills/a2ui-generation/) — a standalone skill you can mount into any agent that supports the Agent Skills mechanism (Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, etc.). It guides the LLM to turn a natural-language query into A2UI protocol messages AGenUI can render directly. Install with npx skills add AGenUI/AGenUI.
• AGenUI Studio — a local, bring-your-own-key workbench (npx agenui-studio) that converts natural language into A2UI protocol with a browser UI for streaming generation, live protocol preview, validation, and one-tap push to the Playground on a real device via QR code. Supports multiple LLM providers (DeepSeek, Qwen, GLM, OpenAI, Gemini, etc.).

Project structure:

• core/ — C++ core engine (parser, differ, layout, Function Call framework)
• platforms/ios|android|harmony/ — native renderers plus bridging layers (Objective-C, JNI, NAPI)
• playground/ — demo apps per platform for development and debugging
• scripts/ — build scripts for each platform

Released under Apache 2.0, it's under active development (current version v1.7.0, released 2026-10-19). Note the changelog is dated in late 2026, so this is a fast-moving project.