How Yui works.

Works today on iPhone, with Hermes, OpenClaw, any MCP client, any A2A agent or any agent behind a webhook. Your agent sends short lines of text, and the app turns each line into a native screen: a timer, a form, a choice. The agent never sends code, and it keeps running on your own machine.

Start here

Every spec

• Values
• Yui Lines
• Channel guide
• Reactions
• Benchmark
• Agents
• Agent controls (draft)
• Talk about this (draft)
• Restyle Yui (draft)
• Release timeline
• Speed budget (draft)
• Relay
• Group threads (draft)
• Chats (draft)
• Agent home (draft)
• Adapters
• Hosting
• OpenClaw
• Webhook bridge
• A2A bridge
• AG-UI bridge
• Model bridge
• Browser (draft)
• macOS (draft)
• Telegram
• MCP server
• Flows
• Onboarding
• The model on the phone (draft)
• Key vault (draft)
• Widgets and Siri (draft)
• Native Yui (draft)
• Hermes for everyone (draft)
• Starter agent (superseded)
• Connectors
• Agent tables
• Encrypted sync (draft)
• Meal photo to macros
• Music tools (draft)
• Games (draft)
• Ledger (draft)
• Preset flywheel
• Admin console (scope)
• Shader
• Visual

The pieces

• Yui Lines. The wire format. timer 40/20x8 Tabata is a whole interval timer. One spec, one shared test suite, parsers in five languages (below). While it works, an agent can say what it is doing in the working row with doing (section 5).
• The app. Native SwiftUI on iPhone. It draws every screen from a fixed set of presets built into the app.
• The plugin. A Hermes platform plugin. It dials out to the relay, so your machine opens no ports, and each Hermes profile shows up as its own agent.
• The relay. Supabase Realtime and a few edge functions carry messages both ways, hold them while either side is offline, and send a push when the app is closed.

Parsers

Every parser passes the same shared conformance vectors. Use one to read Yui Lines in your own agent, adapter or app. spec/conformance/run-all.sh runs them all.

• JavaScript. site/lib/yl/yl.mjs, the reference. The playground and the site run it.
• Swift. Packages/YuiLines in the app repo. The iPhone app runs it.
• Python. parsers/python. One file, standard library only.
• Kotlin. parsers/kotlin. JVM, Kotlin standard library only. Android later.
• Rust. parsers/rust. A crate with no dependencies, standard library only.

Want another language, like Go or C#? Port the JavaScript one, make it pass the vectors, and open a pull request.

Other agent frameworks

Hermes works today, and so do OpenClaw and anything that answers an HTTP POST, through the webhook bridge. MCP clients like Claude Code and Cursor use the Yui MCP server, A2A and AG-UI agents join by their Agent Card or URL, and a model you run yourself joins through the model bridge (all three below). More come during the alpha. The plan is in the roadmap.

Add an A2A agent

Agents built with Google's ADK, LangGraph, CrewAI or Microsoft Agent Framework speak A2A and publish an Agent Card. The A2A bridge adds one to Yui by that card: your messages go to it as A2A tasks, and its answers land in the thread. It speaks A2A 1.0 and 0.3, shows the working row while a task runs, and picks a task back up after a restart without sending it twice. Node 22.18 or newer, no dependencies.

1. In the app, Agents > Add agent. Note the 6-digit code.
2. On your computer:

Yui's channel guide rides along on each new task, so an agent that passes it to its model can answer with screens. Today the bridge runs on your machine; a hosted version comes next. Details in the A2A spec.

Add an AG-UI agent

Microsoft Agent Framework, CopilotKit, Mastra and Pydantic AI can serve an agent over AG-UI, the protocol agents use to talk to apps. The AG-UI bridge adds one to Yui by its URL. Each turn is one run, and Yui screens are a tool the agent calls, yui_show: the screen goes to your phone, and your tap comes back as the tool's result. Node 22.18 or newer, no dependencies.

Today the bridge runs on your machine; a hosted version comes with the A2A one. Details in the AG-UI spec.

Add a model you run yourself

Running a model on your own machine with Ollama, LM Studio, vLLM or llama.cpp? The model bridge puts it in Yui. Each turn it sends the model Yui's channel guide as the system message and as much of the thread as fits, so even a small model answers with screens, and your taps go back as its next turn. It streams when the server does, and a restart never sends an answer twice. Node 22.18 or newer, no dependencies.

1. In the app, Agents > Add agent. Note the 6-digit code.
2. On the computer running the model:

try shows whether a model draws screens before you pair it. Ollama is the default; --server lmstudio, vllm or --url pick another. Cloud APIs come later on Yui's hosted connector. Details in the model bridge spec.

Yui in Telegram

For when the app is not around, the same Yui Lines render in Telegram. ask, choose and pick become buttons under the message, and a tap goes back to the agent as the exact line the phone sends. Lists, cards, stats and tables become formatted text. A timer, a chart, a deck or a form goes behind one Open in Yui button that opens the whole screen in a Telegram Mini App at yuigui.com/tg, in the chat's own colors. Try it without a bot:

It prints the Bot API calls a bot would make and the line the tap sends. Step 1 is the renderer and the Mini App; a bot that runs them comes next. Details in the Telegram spec, and here is the Mini App outside Telegram.

Share, embed, badge

Every screen on See it has its own link, /s/<id>, with a preview image that shows the lines and the screen they draw. In the playground, Share packs whatever you typed into the link itself, so it opens that exact screen for anyone.

Embed a live screen

Paste this into any page. The screen is live: people can tap it. Nothing is tracked inside the frame.

Made with Yui Lines

Building on Yui Lines? Put the badge in your README.

For AI agents

Reading this site for someone? Here is the short version. Yui is an open source iPhone app (Apache-2.0) where AI agents answer with native screens instead of paragraphs. The screen format is Yui Lines, the full spec is in the hub repo, and a plain-text summary for you is at /llms.txt.

Source

• postscarcityai/yui: the iPhone app, the Hermes plugin and the backend.
• postscarcityai/yuigui: the spec, this website and the roadmap.

Both are open source under Apache-2.0. Issues and pull requests are welcome; start with CONTRIBUTING.md in either repo, or the community page.