365 Open Source Plan #002 · Multi-round AI roundtable with history's greatest thinkers
中文 · 繁體中文 · 日本語 · 한국어 · Español · Français · Deutsch · Português · Italiano · Русский · العربية · हिन्दी · Tiếng Việt · ไทย · Türkçe · Indonesia · বাংলা
Put history's greatest minds at one table and let them debate your question.
Legend Talk convenes 2–10 historical or contemporary thinkers into a multi-round debate. Each round, every voice argues from its own framework; a moderator then maps the disagreements and opens the next round. Socrates presses Munger's assumptions while Nietzsche challenges them both.
Three ways in:
- Ask a question — type a topic and AI assembles a 3–5 thinker panel built for productive tension.
- Set the table — hand-pick 2–10 thinkers yourself, or roll 5 at random.
- Consult one mind — go 1-on-1 with any of 161 thinkers, each reasoning through their own framework, not generic AI roleplay.
Demo: talk.newzone.top — 18 languages · free · local-first · no sign-up.
| Home page | Chat view |
|---|---|
 |
 |
Auto roundtable — enter a topic in the input bar on the home page. AI picks 3–5 thinkers whose views genuinely conflict and starts the debate immediately, no character picking needed.
Manual roundtable — click + on 2–10 character cards to build a lineup. A floating bar shows your picks:
- Click an avatar to remove it
- Start Discussion to launch
- Copy lineup link to share the exact lineup as a URL
Or hit 🎲 Random (top-right) to start instantly with 5 random thinkers.
Featured templates — 6 curated lineups whose perspectives genuinely clash (e.g. AI & Tech: Karpathy vs Ilya vs Feynman vs Taleb vs Paul Graham). One click to start, each with 3 suggested topics.
1-on-1 chat — click Chat on any character card for a private conversation in that thinker's voice and framework.
Topic suggestions — before your first message, every mode proposes topics to get you started: 3 curated questions per template (in all 18 languages), or 1 question drawn from each selected thinker in a manual lineup (updates as you add or remove people).
You sit at the head of the table as the chair — the debate runs on your terms.
- Moderator — after each round, an AI moderator synthesizes it: groups claims by idea, names an angle the round left untouched, and poses an open question for the next one.
- Refocus mid-debate — send a message during a roundtable to redirect it. Instead of auto-running, an editable focus card appears; it stacks on top of any prior focus so earlier steers aren't lost. Refine it, then Start the next rounds anchored on it.
- Apply & retry — edit any message and regenerate from that point. Chair steers carry a focus snapshot, so a retry rebuilds from the exact focus state that was active when the message was sent.
- Set rounds — choose how many rounds the thinkers debate before pausing, and Continue to add more once they finish.
- Add or remove participants anytime — turn a 1-on-1 into a roundtable, or the reverse.
- Stop — cancel generation mid-stream; whatever was already written is kept.
- Branch — fork a new conversation from any message, carrying the prior context with it.
- Summarize — one-click AI summary extracting the core viewpoints and disagreements.
- Search — find anything across all conversations by title, thinker name, or message content.
- Favorites — star your most-used thinkers for quick access.
- Share chat — generate a URL containing the full conversation.
- Export / Import — save as Markdown or JSON and restore from JSON (in Settings), or generate share cards via json2card (set the API endpoint in Settings).
- Settings sync — move your setup to another device via URL; API keys are AES-encrypted.
161 preset thinkers across 15 domains, sorted by fame — type any name to create a custom character on the fly.
Models — set the thinking level (off / low / medium / high), enter a custom model ID, or connect any OpenAI-compatible API as a custom provider. Default: DeepSeek V4 Flash.
Platform — 18 languages · dark mode · responsive · local-first (IndexedDB + localStorage dual-write, works in WeChat and restricted WebViews) · zero CDN (fonts self-hosted and bundled, so it works offline and behind firewalls).
Start a conversation straight from a URL:
- By name:
/#/chat?chars=苏格拉底,孔子or/#/chat?chars=Socrates,Confucius - By ID:
/#/chat?chars=socrates,confucius - By category:
/#/chat?category=philosophy(roundtable of every thinker in that category, capped at 10) - Single chat:
/#/chat?chars=socrates - Custom names:
/#/chat?chars=Ada Lovelace,Linus Torvalds(unrecognized names become custom characters)
Categories: philosophy, strategy, business, finance, history, sociology, psychology, science, literature, art, economics, politics, technology, religion, education.
The Copy lineup link (participants bar) and Copy category link (category filter) buttons generate these URLs from the UI.
Language routing — prefix the URL with a language to set the UI language, e.g. /#/ja/chat, /#/ko/chat?chars=socrates, or use ?lang=zh. All 18 languages supported.
24 providers out of the box — international, China-based, and aggregators:
| Provider | Models |
|---|---|
| OpenAI | GPT-5.5, GPT-5.4, GPT-5.4 Mini |
| Anthropic | Claude Opus 4.7, Claude Sonnet 4.6, Claude Haiku 4.5 |
| Google Gemini | Gemini 3.1 Pro, Gemini 3.5 Flash |
| xAI Grok | Grok 4.3, Grok 4.20 series |
| Mistral / Cohere | Mistral Medium 3.5 / Large 3, Command A series |
| DeepSeek | DeepSeek V4 Flash, V4 Pro |
| Moonshot / Kimi | Kimi K2.6, K2.5 |
| Zhipu GLM | GLM-5.1, GLM-5, GLM-4.7 series |
| MiniMax / Hunyuan / Qianfan / MiMo | MiniMax M2.7, Hunyuan 2.0, ERNIE 5.1, MiMo V2.5 |
| Volcengine Coding Plan | Doubao Seed 2.0, Kimi K2.5, GLM-4.7, DeepSeek V4 |
| Alibaba Bailian Coding Plan | Qwen 3.6 Max/Plus/Flash, Kimi K2.5, GLM-5 |
| Aggregators | OpenRouter, SiliconFlow, Groq, Cerebras, Together, Fireworks, Perplexity, NVIDIA NIM, GitHub Models |
Every provider accepts custom model IDs, and the Custom option connects any OpenAI-compatible API.
npm install
npm run devOpen http://localhost:5173, go to Settings, enter your API key, and start chatting. If you hit a CORS error, the app offers to enable a public proxy with one click.
Some providers block direct browser requests. The CORS proxy is configured per provider in Settings — toggle it on. A public proxy (https://cors.api2026.workers.dev) is used by default.
To run your own, deploy a Cloudflare Worker with this code:
Worker code
export default {
async fetch(request) {
const url = new URL(request.url);
const targetUrl = url.pathname.slice(1) + url.search;
if (!targetUrl || !targetUrl.startsWith('https://')) {
return new Response('Usage: /https://target-api.com/path', { status: 400 });
}
if (request.method === 'OPTIONS') {
return new Response(null, {
headers: {
'Access-Control-Allow-Origin': '*',
'Access-Control-Allow-Methods': 'GET, POST, PUT, DELETE, OPTIONS',
'Access-Control-Allow-Headers': '*',
'Access-Control-Max-Age': '86400',
},
});
}
const response = await fetch(targetUrl, {
method: request.method,
headers: request.headers,
body: request.body,
});
const newResponse = new Response(response.body, response);
newResponse.headers.set('Access-Control-Allow-Origin', '*');
return newResponse;
},
};src/
adapters/ # LLM API adapters (OpenAI-compatible, Anthropic)
characters/ # Character presets and custom character generation
components/ # React components
hooks/ # useChat, useRoundtable
i18n/ # Internationalization
stores/ # Zustand state management
utils/ # Prompt building, export, compression, storage
types.ts # Type definitions
Stack: React 19 · antd 6 (CSS-variables theme, deeply customized) · Vite · Tailwind CSS v4 · Zustand · i18next · React Router · TypeScript
| Command | Description |
|---|---|
npm run dev |
Start dev server |
npm run build |
Type-check and build for production |
npm run test |
Run tests |
npm run preview |
Preview production build |
Build and host the dist/ folder on any static host (Vercel, Netlify, GitHub Pages, …):
npm run buildRouting is hash-based (/#/chat/..., /#/ja/chat/...), so no server-side routing config is needed.
This is project #002 of the 365 Open Source Plan — one person + AI, 300+ open source projects in a year. Submit your idea →
MIT