Your AI, your interface

Use Claude or ChatGPT in a screen you design yourself, just by saying how you want it. In Unchore you connect your own Claude or ChatGPT plan (or a Gemini, Grok, DeepSeek or Kimi key). Everyone starts with the same screen. Say "bigger text, menu on the right" and only your screen changes. Same AI, a different interface for every person.

This repo is the pattern behind that feature, boiled down to one dependency-free file you can read in five minutes.

▶ Try the live demo (no signup, no AI key)

The full app is open source too: smilemino/unchore-ai (AGPL-3.0).

Why

AI chat apps ship one layout for millions of people. A designer picks where the menu goes, and everyone lives with it, whether they use the AI for code, sales or homework. Now that an AI can write a screen from a sentence, that choice can belong to each user:

• "Bigger text, and put the menu on the right."
• "Show today's to-dos above the chat."
• "A focus timer next to the chat."

The hard part is not getting an AI to write the screen. It is letting AI-written code run inside a logged-in app without it reading other people's data, phoning home, or locking the user out. Ink & Switch calls this direction malleable software. This repo is one small, practical piece of it.

How it stays safe

sandbox.js (about 130 lines) does four things:

	What happens	What it stops
1	The screen is drawn inside an iframe with sandbox="allow-scripts" and no allow-same-origin	Reading your app's cookies, storage or page
2	The screen's logic runs in a Web Worker started inside that iframe, under connect-src 'none'	fetch, WebSocket, importScripts to anywhere
3	A small trusted renderer keeps an allow-list of tags and attributes	<script>, <img>, <a>, <form>, every on* handler
4	A new screen is test-run hidden for 1.5 s, and a watchdog pings it every 250 ms	Broken or frozen screens replacing a working one

The "back to default" button lives outside the sandbox, so a bad screen can never hide the way out.

The AI-written code only gets four calls:

ui.render(html, css)      // draw (sanitized)
ui.onAction(fn)           // clicks and changes on [data-action] elements
ui.load(key)              // this user's own saved data
ui.save(key, value)

import { mountScreen } from './sandbox.js';

const result = await mountScreen(container, aiWrittenCode, {
  store: { load: k => db.get(userId, k), save: (k, v) => db.set(userId, k, v) },
  onFail: () => showDefaultScreen(),
});
if (!result.ok) keepCurrentScreen(result.reason); // 'froze' | 'crashed' | 'never drew'

What the demo checks

test/demo_test.py drives the demo in headless Chromium. 13 of 13 checks pass:

• a typed request moves the menu to the right and makes the text bigger
• a to-do added inside the user's screen survives a reload
• an endless loop is rejected in its test run, the old screen stays, the page stays responsive
• fetch, WebSocket, importScripts, cookies and localStorage are all blocked, and no request reaches the outside site
• an injected <script> and onclick never run
• reset brings back the default screen, and a second visitor still sees the default

pip install playwright && playwright install chromium
python test/demo_test.py

In Unchore

In the demo the edits are pre-written so it runs without an AI. In Unchore you just say it in the chat, with whichever AI you connected: your Claude plan, your ChatGPT plan, or a key for Gemini, Grok, DeepSeek or Kimi. The AI writes the screen, it is test-run, saved to your account only, and you can go back to the default screen at any time.

Unchore is a personal AI: ask anything, and if it repeats, Unchore offers to automate it. Humans do the creative work. AI does the chores. → unchore.ai

Limits

• Only tested in Chromium so far. Firefox and Safari are not tested yet.
• This is a pattern, not a security audit. CSS can still make a screen ugly or hard to read; that is why the way back sits outside the sandbox.
• The demo stores each visitor's screen in localStorage. A real app should store it per user on the server.

한국어 요약

클로드·챗GPT를 내가 원하는 화면으로 쓰세요. 언초어에 내 클로드나 챗GPT 요금제(또는 제미나이·그록·딥시크·키미 열쇠)를 연결하고 화면을 말로 바꾸면 됩니다. 처음에는 모두 같은 화면으로 시작하고, 회원이 «글자 크게, 메뉴는 오른쪽에»처럼 말하면 그 사람 화면만 바뀝니다. AI가 만든 화면 코드는 인터넷에 접속할 수 없고 쿠키나 저장 공간도 읽지 못하는 격리된 공간에서만 실행됩니다. 먼저 보이지 않게 시험해 보고 통과한 화면만 적용되며, 기본 화면으로 돌아가는 단추는 늘 격리된 공간 밖에 있습니다. 언초어에서 실제로 쓸 수 있습니다.

MIT License

Claude, ChatGPT, Gemini, Grok, DeepSeek and Kimi are trademarks of their owners. Unchore is not affiliated with them; it connects to the plan or key you already have.